Collaborative filtering recommendation method based on attribute knowledge introduction

Abstract

The present invention belongs to the field of recommendation technology and discloses a collaborative filtering recommendation method based on the introduction of attribute knowledge, including: constructing an attribute system based on direct and indirect user information; feature selection based on a logistic regression model with an L1 penalty term; fusing user behavior vectors and user attribute knowledge vectors based on dual-vector splicing; and integrating collaborative filtering, neighbor calculation based on user attribute knowledge, and fusing user attribute knowledge with behavior vectors to form a triple hybrid recommendation, wherein the triple hybrid recommendation utilizes attribute information and integrates user attributes, behavior data, and collaborative filtering recommendations through weighted voting. The present invention constructs and introduces attribute information, integrates behavior data information, and establishes a triple hybrid recommendation model based on embedded feature selection and vector splicing similarity calculation to combine user attributes and collaborative filtering, thereby improving the recommendation effect.

Description

Technical Field

[0001] The present invention belongs to the technical field of recommendation, and in particular relates to a collaborative filtering recommendation method based on the introduction of attribute knowledge. Background Art

[0002] The rapid adoption of e-commerce, video services, and news services has continuously increased convenience for consumers. However, accurately matching products, news, and videos with potential consumers is becoming increasingly difficult. Potential consumers need technology that can help them find products that better suit their interests. The goal of information recommendation technology is to bring various items (including news, products, videos, etc.) to the right users while maintaining a high matching accuracy rate. However, when using recommendation technology to overcome the problem of "information overload," several issues that hinder the improvement of recommendation accuracy are becoming increasingly prominent.

[0003] Recommendation technologies often face scenarios involving large amounts of data, but insufficient user-related data to support recommendations. Data sparsity not only increases computational complexity but also reduces the effectiveness of recommendations. Furthermore, user rating patterns vary significantly from user to user, leading to reduced accuracy in both user similarity calculations and the resulting recommendations when using rating patterns to simulate user interests.

[0004] Improvements to recommendation algorithms based on collaborative filtering require new methods for representing user interests and calculating user similarity.

[0005] Collaborative filtering is a promising approach for meeting recommendation needs. It considers user behavior as a representation of interest and categorizes users based on their historical ratings. Therefore, user grouping also reflects user preferences to a certain extent.

[0006] Traditional collaborative filtering methods have spawned two types of algorithms: user-based collaborative filtering (UBCF) and item-based collaborative filtering (IBCF). The assumption behind user-based collaborative filtering is that predicted ratings can intuitively indicate a user's interest level, enabling appropriate recommendations. Item-based collaborative filtering calculates the similarity between items, which is crucial for subsequent recommendation calculations. Item-based collaborative filtering is also widely used, particularly with semi-structured or unstructured data, when extracting item attributes is difficult.

[0007] With the rapid and widespread growth of online users and the availability of diverse information, recommendation algorithms are increasingly experiencing issues such as large errors in understanding user interests, deviations from user interests, and topic drift. These shortcomings are hindering the effectiveness of recommendations and have become urgent challenges in the recommendation field. Summary of the Invention

[0008] In view of this, the present invention proposes a collaborative filtering recommendation method based on attribute knowledge introduction, which includes the following steps:

[0009] Define the user's direct and indirect information, and build the user's attribute system based on this; the collection of all attributes generates the entire dimension of the user's attributes;

[0010] In order to screen key features with large information content and high discrimination, a logistic regression model with L1 penalty term is used for feature selection;

[0011] Vectorize user attribute knowledge and behavior data separately, and combine the user behavior vector with the user attribute knowledge vector to form a recommendation that integrates attributes and behaviors.

[0012] A triple hybrid recommendation is constructed based on the attribute knowledge embedding selection method with L1 penalty term and user attribute knowledge embedding collaborative filtering. The triple hybrid recommendation enriches the user's interest preference expression by utilizing attribute information, improves the accuracy of similarity calculation between users, and integrates user attributes and collaborative filtering through weighted voting.

[0013] Furthermore, indirect user attributes are defined and associated with project information associated with user behavior. If a user rates an item highly, the attributes of the item will be highlighted in the user's corresponding attribute dimension. An item may have multiple attributes, so a rating may highlight multiple dimensions in the user's attribute knowledge space. If a user has never rated any item with the same attributes, or the rating is not good enough, the element values ​​corresponding to those item attributes in the user attribute vector will be 0. Whenever a user rates an item with one or more attributes and the rating is good, the user attribute dimensions corresponding to these item attributes will increase.

[0014] Furthermore, the construction of the attribute system based on the user's direct information and indirect information includes:

[0015] ADV represents the element value in the user attribute knowledge dimension. ADV is the value calculated in each round, and the following is true:

[0016]

[0017] j represents the jth item, and k represents the attribute of the item;

[0018]

[0019] i represents the user i under investigation, z represents the zth user attribute knowledge dimension, n and p are the total number of projects and the total number of project attributes, respectively;

[0020] For user direct attributes, the feature space is generated directly according to the direct attributes, and the direct attribute value advd is quantified according to whether it has the attributes on the corresponding dimension:

[0021]

[0022] x represents the xth attribute.

[0023] Furthermore, from the attribute set to the formation of attribute knowledge, logistic regression is used to select features, filter out attributes with low discrimination, and select a small number of features that can best reflect user preferences to form knowledge based on user attributes. The logistic regression formula is as follows:

[0024]

[0025] z=a1x1+a2x2+…+a n x n +b

[0026] x1,x2,x n are the input features to be selected, a1, a2, a n is the weight, z is the output result after calculation, and b is a constant.

[0027] Furthermore, the loss function is set for the logistic regression function, and the log-likelihood loss function is used as follows:

[0028]

[0029] When faced with the problem of solving the extreme value of an unconstrained multivariate function, the gradient descent method is used to iteratively obtain the model parameters corresponding to the minimized loss function.

[0030] Furthermore, in order to select a small number of attribute features with high discrimination in the user attribute system, a penalty mechanism is added when solving the regression coefficient. This not only reduces the complexity of the model and prevents overfitting in the training process, thereby reducing the generalization ability of the model, but also plays a role in feature screening. Therefore, the L1 regularization coefficient is selected as the penalty term. L1 is the sum of the absolute values ​​of each parameter in the model. The formula is as follows:

[0031]

[0032] Where A is the regression parameter a i The column vector is composed of λ||A||1, λ is the L1 regularization term, λ is the adjustable parameter of the L1 regularization term, y i is an estimated value, x i is the sample eigenvalue, a i is the regression coefficient, and N is the total number of samples.

[0033] Furthermore, the fusion of user attribute knowledge vector and behavior vector based on dual vector concatenation includes:

[0034] The first step in using user attribute knowledge is to normalize the user attribute knowledge vector so that the attribute values ​​do not vary significantly between different users. The normalization formula is as follows:

[0035]

[0036] uakv'(i) represents the normalized value of user i, uakv(i) represents its original value, and k is the dimension of the user attribute knowledge vector.

[0037] Furthermore, a triple hybrid recommendation based on user attribute knowledge embedding collaborative filtering is proposed, including: the similarity between vectors is achieved by using cosine similarity or Pearson correlation coefficient.

[0038] Furthermore, when considering the ratings of each most similar user on an item, the target user's attitude towards the item will be estimated by the average of the ratings of all the most similar users:

[0039]

[0040] R(i, j) represents the rating data from user i to item j, and R(w, j) represents the rating data from user w to item j, where w is one of the users most similar to user i, and the total number is k;

[0041] By comprehensively considering the fusion method of traditional collaborative filtering, user attribute knowledge and user attribute knowledge embedded in collaborative filtering, a triple hybrid recommendation model is constructed based on the inter-model recommendation result voting method. It can integrate user attribute information and scoring mode to improve similar user mining;

[0042] R = αR CF +βR uak +(1-α-β)R vp

[0043] R represents the final recommendation for the target user, R CF represents the recommendation through traditional collaborative filtering algorithm, R uak represents the recommendation through the user attribute knowledge embedding method, R vp It is a recommendation based on user fusion vectors, and α and β are adjustable parameters in the triple hybrid recommendation model.

[0044] The beneficial effects of the present invention are as follows:

[0045] This paper studies the role of attribute knowledge in improving collaborative filtering recommendations. It defines direct and indirect user information. By constructing a user attribute system and embedding user attribute knowledge, it proposes a triple hybrid recommendation algorithm: an attribute knowledge embedding and selection method based on an L1 penalty term, and a method that embeds user attribute knowledge into collaborative filtering. By leveraging attribute information, the similarity between users is increased. This hybrid model algorithm combines user attributes with collaborative filtering through weighted voting, enhancing recommendation effectiveness. Experiments have verified the effectiveness of this invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of the concatenation and fusion of user attribute vectors and behavior vectors;

[0047] Figure 2 Logic process diagram of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be further described below with reference to the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention fall within the scope of protection of the present invention.

[0049] User attributes reflect their unique characteristics. Leveraging the information contained in these attributes to generate user knowledge will help improve the accuracy of user recommendations. User attributes can be direct, such as the identity, location, and interests provided by users during registration, or indirect, such as user preferences derived from analyzing user behavior. By combining two types of user attributes to construct a user attribute system, we select features from these attributes using a logistic regression model with an L1 penalty, mining the information within these attributes to generate highly condensed user attribute knowledge that reflects key user characteristics.

[0050] Construction of attribute system based on user direct and indirect information

[0051] The user attribute system is set as a collection of direct and indirect user attributes. To build a rich user attribute system, it is necessary to introduce indirect attribute information based on user behavior data, forming a representation of user interests expressed in user behavior, which together with the user's direct attributes constitutes a complete user attribute system.

[0052] The specific process of indirect attribute information introduction is as follows: Rating data establishes a connection between users and items, and ratings also introduce additional information related to items for a specific user. For example, a rating from user u for an item labeled "belonging to the computer field" means that user u is likely interested in information related to the computer field. Therefore, this can be considered as the knowledge that the user has gained from item attributes. As more rating data is considered, the user's indirect attribute knowledge becomes increasingly comprehensive, and it can largely represent the user's interests.

[0053] In this scenario, the collection of all item attributes forms the entire dimension of user attribute knowledge. If a user rates an item highly, the attributes associated with that item will be highlighted in the user's corresponding attribute dimension. It's worth noting that an item may have multiple attributes, so a single rating may highlight multiple dimensions in the user's attribute knowledge space. In other words, if a user has never rated any items with the same attributes, or if their ratings are suboptimal, the values ​​of the elements in the user attribute vector corresponding to those item attributes will be 0. Whenever a user rates an item with one or more attributes and gives a favorable rating, the user attribute dimension corresponding to those item attributes will increase.

[0054] Specifically, if a user's rating of an item is higher than the average of all its ratings, then the rating is good enough, otherwise it is not. Let ADV represent the element value in the user attribute knowledge dimension, and ADV is the value calculated in each round, then we have the following:

[0055]

[0056] j represents the jth item, and k represents the attribute of the item.

[0057]

[0058] i represents the user i under investigation, z represents the zth user attribute knowledge dimension, n and p are the total number of projects and the total number of project attributes, respectively.

[0059] For user direct attributes, the feature space is generated directly according to the direct attributes, and the direct attribute value advd is quantified according to whether it has the attributes on the corresponding dimension:

[0060]

[0061] x represents the xth attribute.

[0062] In this way, we have constructed the user's attribute system, as well as the representation of direct and indirect attributes and their vectorization process, and set the quantization and accumulation rules of element values ​​in the vector.

[0063] Feature selection based on logistic regression model with L1 penalty

[0064] The construction of a user attribute system maximizes the collection of user characteristics and can model all potential factors that directly or indirectly reflect user interests. However, from attribute collection to the formation of attribute knowledge, a feature selection process is required to filter out those attributes with low discriminability and select a small number of features that best reflect user preferences to form knowledge based on user attributes.

[0065] Logistic regression is a model that achieves classification by establishing a regression formula for the boundary line of the data. In order to handle the binary classification problem, the Sigmoid function is used as the activation function. In this way, the processing to be carried out by the logistic regression model is the following steps: the eigenvalue is multiplied by the regression coefficient, and the product is added as the input of the Sigmoid function, which corresponds to a function value ranging from 0 to 1. According to the definition of binary classification, function values ​​greater than 0.5 are classified into one category, and the others are classified into the other category. The logistic regression formula is as follows:

[0066]

[0067] z=a1x1+a2x2+…+a n x n +b

[0068] In order to obtain the set of regression coefficients, it is necessary to use the training data set through learning. Therefore, the loss function is set for the logistic regression function, and the log-likelihood loss function is used.

[0069]

[0070] Faced with this unconstrained multivariate function extreme value solution problem, the gradient descent method is used to iteratively obtain the model parameters corresponding to the minimized loss function.

[0071] Specifically in the scenario of this method, the goal is to select a small number of attribute features with high discrimination from the user attribute system. When solving the regression coefficient, a penalty mechanism needs to be added to reduce the complexity of the model, prevent overfitting during the training process and thus reduce the generalization ability of the model, and also play a role in feature screening. Therefore, the L1 regularization coefficient is selected as the penalty term. L1 is the sum of the absolute values ​​of each parameter in the model. The formula is as follows:

[0072]

[0073] Where A is the regression parameter a i The column vector is composed of λ||A||1, λ is the L1 regularization term, λ is the adjustable parameter of the L1 regularization term, yi is an estimated value, x i is the sample eigenvalue, a i is the regression coefficient, and N is the total number of samples.

[0074] After feature selection using a logistic regression model with an L1 penalty, we select attributes with sufficient information and high discrimination, which have non-zero regression parameters. Other regression parameters with a value of 0 represent attributes that can be filtered out. In this way, we use an embedded feature selection method to select a subset of attributes from the user attribute system that reflect user characteristics and preferences, forming the attribute knowledge that reflects user interests.

[0075] Embedding of user attribute knowledge

[0076] The vector representation of user interest preferences includes user attribute knowledge vector and user behavior vector.

[0077] Fusion of user attribute knowledge vector and behavior vector based on dual vector concatenation

[0078] The first step in using user attribute knowledge is to normalize the user attribute knowledge vector so that the attribute values ​​do not vary significantly between different users. The normalization formula is as follows:

[0079]

[0080] uakv'(i) represents the normalized value of user i, uakv(i) represents its original value, and k is the dimension of the user attribute knowledge vector.

[0081] In collaborative filtering, user behavior data on items represents their interest. Therefore, a behavior vector is generated for each user. The dimension of the vector is the number of items, and each element in the vector represents the user's level of interest in the item along that dimension. The user behavior vector is a model of user interest preferences derived by quantifying user behaviors such as clicks, favorites, purchases, and reviews. This vector is generated offline for all users by fully processing their historical behavior data.

[0082] At this point, for each user, an attribute knowledge vector and a behavior vector have been generated. These two vectors each describe the user's interests and preferences from different perspectives. To fully utilize user information and recommend items that more accurately and comprehensively reflect their needs, it is necessary to comprehensively utilize these two vectors.

[0083] Vector concatenation is a method for processing multiple fixed-length vectors of the same subject. By concatenating two or more vectors in a specific order to form a single vector, it integrates multi-dimensional information, expands the feature space, and embeds it into a single vector representation. Here, we place the user attribute knowledge vector first and the user behavior vector last, forming a fusion of user attribute knowledge and behavior data.

[0084] Triple hybrid recommendation based on user attribute knowledge embedding collaborative filtering

[0085] The similarity between vectors can be achieved by using cosine similarity or Pearson correlation coefficient. The corresponding formulas are shown below.

[0086]

[0087]

[0088] Cosine similarity is the normalized dot product result, while the Pearson correlation coefficient is the decentralized dot product. Both provide a value representing the similarity between vectors, reflecting the likelihood that users share common interests while also exhibiting certain differences. Cosine similarity is more suitable for scenarios with large differences in rating dimensions, such as calculating the similarity of user fusion vectors. Therefore, when calculating user attribute knowledge vectors and user behavior vectors, we will use cosine similarity to obtain a similarity that better reflects the user's own characteristics, while using the Pearson correlation coefficient when calculating user fusion vectors.

[0089] Generally speaking, the most similar users are likely to have very similar attribute preferences and rating patterns, and can serve as a reference for each other. Therefore, items rated by the most similar users are likely to be attractive to the target user. For the target user, the item ratings of the most similar users are also highly referenceable.

[0090] Traditional collaborative filtering algorithms mine rating data and identify rating patterns to find similar users with shared interests. However, rating data and rating patterns are indirect reflections of user interests, and using them to predict recommendations for target users is not always accurate. In contrast, attribute knowledge directly reflects user interests. Therefore, using this information to find users with shared interests can ensure reasonable results. This means that after obtaining similarities between users, the preferences of the most similar users can be used to filter items that the target user may be interested in. If many of the most similar users have rated an item highly, the target user is likely to rate it, and the rating is likely to be high. When considering the ratings of each most similar user for an item, the target user's attitude towards the item is estimated by averaging the ratings of all the most similar users.

[0091]

[0092] R(i, j) represents the rating data from user i to item j, and R(w, j) represents the rating data from user w to item j, where w is one of the users most similar to user i, and the total number is k.

[0093] Taking into account the fusion method of traditional collaborative filtering, user attribute knowledge and user attribute knowledge embedded in collaborative filtering, a triple hybrid recommendation model is constructed based on the inter-model recommendation result voting method, which can integrate user attribute information and scoring mode to improve similar user mining.

[0094] R = αR CF +βR uak +(1-α-β)R vp

[0095] R represents the final recommendation for the target user, R CF represents the recommendation through traditional collaborative filtering algorithm, R uak represents the recommendation through the user attribute knowledge embedding method, R vp Recommendation is based on user fusion vectors. α and β are adjustable parameters in the triple hybrid recommendation model. Their effects can be studied in the experimental section and the optimal values ​​can be selected.

[0096] The beneficial effects of the present invention are as follows:

[0097] This paper studies the construction of a user attribute system and the embedded selection of user attribute knowledge. It proposes a triple hybrid recommendation algorithm that embeds attribute knowledge selection using an L1 penalty term and embeds user attribute knowledge into collaborative filtering. By leveraging attribute information, the similarity between users is improved. This hybrid model algorithm combines user attributes with collaborative filtering through weighted voting, enhancing recommendation effectiveness. Experiments have verified the effectiveness of this invention.

[0098] As used herein, the word "preferred" is intended to serve as an example, instance, or illustration. Any aspect or design described herein as "preferred" is not necessarily to be construed as advantageous over other aspects or designs. Rather, the use of the word "preferred" is intended to present concepts in a concrete manner. As used in this application, the term "or" is intended to mean an inclusive "or" rather than an exclusive "or." That is, unless otherwise specified or clear from the context, "X employs A or B" is intended to mean any of the naturally inclusive permutations. That is, if X employs A; X employs B; or X employs both A and B, then "X employs A or B" is satisfied in any of the foregoing examples.

[0099] Moreover, although the present disclosure has been shown and described with respect to one or implementation, those skilled in the art will think of equivalent variations and modifications based on reading and understanding of this specification and the accompanying drawings. The present disclosure includes all such modifications and variations and is limited only by the scope of the appended claims. In particular, with respect to the various functions performed by the above-mentioned components (such as elements, etc.), the terms used to describe such components are intended to correspond to any component (unless otherwise indicated) that performs the specified function of the component (such as it is functionally equivalent), even if structurally different from the disclosed structure that performs the function in the exemplary implementation of the present disclosure shown herein. In addition, although the specific features of the present disclosure have been disclosed with respect to only one of several implementations, such features can be combined with one or other features of other implementations that can be desired and advantageous for a given or specific application. Moreover, insofar as the terms "including", "having", "containing" or their variations are used in specific embodiments or claims, such terms are intended to be included in a manner similar to the term "comprising".

[0100] The functional units in the embodiments of the present invention may be integrated into a single processing module, or each unit may exist physically separately, or multiple or more units may be integrated into a single module. The aforementioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium. The aforementioned storage medium may be a read-only memory, a magnetic disk, or an optical disk, etc. The aforementioned devices or systems may execute the storage method in the corresponding method embodiment.

[0101] In summary, the above embodiment is one implementation method of the present invention, but the implementation method of the present invention is not limited to the described embodiment. Any other changes, modifications, substitutions, combinations, and simplifications that deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. The collaborative filtering recommendation method based on attribute knowledge is characterized by: The following steps are involved: Define the user's direct and indirect information, and build the user's attribute system based on this; the collection of all attributes generates the entire dimension of the user's attributes; In order to screen key features with large information content and high discrimination, a logistic regression model with L1 penalty term is used for feature selection; Vectorize user attribute knowledge and behavior data separately, and combine the user behavior vector with the user attribute knowledge vector to form a recommendation that integrates attributes and behaviors. A triple hybrid recommendation method based on attribute knowledge embedding selection method of L1 penalty term and user attribute knowledge embedding collaborative filtering is constructed. The triple hybrid recommendation enriches the user's interest preference expression by utilizing attribute information, improves the accuracy of similarity calculation between users, and integrates user attributes and collaborative filtering through weighted voting. Define user indirect attributes and use the project information associated with user behavior to associate them. If a user gives a high rating to an item, the attributes of the item will be highlighted in the user's corresponding attribute dimension. An item may have multiple attributes, so a rating may highlight multiple dimensions in the user's attribute knowledge space. If a user has never rated any item with the same attributes, or the rating is not good enough, the element value corresponding to those item attributes in the user attribute vector will be 0. Whenever a user rates an item with one or more attributes and the rating is good, the user attribute dimension corresponding to these item attributes will increase. The attribute system based on the user's direct and indirect information is constructed, including: ADV represents the element value in the user attribute knowledge dimension. ADV is the value calculated in each round, and the following is true: j represents the jth item, and k represents the attribute of the item; i represents the user i under investigation, z represents the zth user attribute knowledge dimension, n and p are the total number of projects and the total number of project attributes, respectively; For user direct attributes, the feature space is generated directly according to the direct attributes, and the direct attribute value advd is quantified according to whether it has the attributes on the corresponding dimension: x represents the xth attribute.

2. The collaborative filtering recommendation method based on attribute knowledge introduction according to claim 1 is characterized in that: In order to form attribute knowledge from attribute sets, we use logistic regression for feature selection, filter out attributes with low discrimination, select a small number of features that can best reflect user preferences, and form knowledge based on user attributes. The logistic regression formula is as follows: z=a1x1+a2x2+…+a n x n +b x1,x2,x n are the input features to be selected, a1, a2, a n is the weight, z is the output result after calculation, and b is a constant.

3. The collaborative filtering recommendation method based on attribute knowledge introduction according to claim 2 is characterized in that: Set the loss function for the logistic regression function and use the log-likelihood loss function as follows: When faced with the problem of solving the extreme value of an unconstrained multivariate function, the gradient descent method is used to iteratively obtain the model parameters corresponding to the minimized loss function.

4. The collaborative filtering recommendation method based on attribute knowledge introduction according to claim 1 is characterized in that: In order to select a small number of attribute features with high discrimination in the user attribute system, a penalty mechanism is added when solving the regression coefficient. This not only reduces the complexity of the model and prevents overfitting in the training process, thereby reducing the generalization ability of the model, but also plays a role in feature screening. Therefore, the L1 regularization coefficient is selected as the penalty term. L1 is the sum of the absolute values ​​of each parameter in the model. The formula is as follows: Where A is the regression parameter a i The column vector is composed of λ||A||1, λ is the L1 regularization term, λ is the adjustable parameter of the L1 regularization term, y i is an estimated value, x i is the sample eigenvalue, a i is the regression coefficient, and N is the total number of samples.

5. The collaborative filtering recommendation method based on attribute knowledge introduction according to claim 1 is characterized in that: The fusion of user attribute knowledge vector and behavior vector based on dual vector splicing includes: Normalize the user attribute knowledge vector so that the attribute values ​​do not vary too much between different users. The normalization formula is as follows: uakv'(i) represents the normalized value of user i, uakv(i) represents its original value, and k is the dimension of the user attribute knowledge vector; Generate an attribute knowledge vector and a behavior vector for each user. The behavior vector is obtained by collecting and quantifying the user's click, favorite, purchase, and evaluation behaviors to obtain the user's interest preferences. The dimension of the behavior vector is the number of all items, and each element value represents the user's degree of interest in the items in the corresponding dimension. The user attribute knowledge vector is set in front and the user behavior vector is set in the back to form a fusion of user attribute knowledge and behavior data, expand the feature space, and embed them into the same vector expression.

6. The collaborative filtering recommendation method based on attribute knowledge introduction according to claim 1 is characterized in that: Triple hybrid recommendation based on user attribute knowledge embedding collaborative filtering, including: the similarity between vectors is achieved by using cosine similarity or Pearson correlation coefficient.

7. The collaborative filtering recommendation method based on attribute knowledge introduction according to claim 1 is characterized in that: When considering the ratings of each most similar user on an item, the target user’s attitude towards the item will be estimated by the average of all the most similar users’ ratings: R(i, j) represents the rating data from user i to item j, and R(w, j) represents the rating data from user w to item j, where w is one of the users most similar to user i, and the total number is k; By comprehensively considering the fusion method of traditional collaborative filtering, user attribute knowledge and user attribute knowledge embedded in collaborative filtering, a triple hybrid recommendation model is constructed based on the inter-model recommendation result voting method. It can integrate user attribute information and scoring mode to improve similar user mining; R=αR CF +βR uak +(1-α-β)R vp R represents the final recommendation for the target user, R CF represents the recommendation through traditional collaborative filtering algorithm, R uak represents the recommendation through the user attribute knowledge embedding method, R vp It is a recommendation based on user fusion vectors, and α and β are adjustable parameters in the triple hybrid recommendation model.

Images