A concept-based recommendation method and device based on user feedback and search content

By constructing a concept grid recommendation method based on user record feedback and search content, the problems of data sparsity, cold start and noise in IoT recommendation systems are solved, the recommendation accuracy and quality are improved, and it is adapted to dynamic and heterogeneous data environments.

CN116204721BActive Publication Date: 2026-11-17NANJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310228763.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-03-10
Publication Date
2026-11-17
Estimated Expiration
2043-03-10

AI Technical Summary

Technical Problem

Traditional recommendation systems fail to effectively utilize dynamic and heterogeneous IoT data when building IoT recommendation systems, resulting in data sparsity, cold start, and noise issues, leading to low recommendation accuracy.

Method used

By constructing a concept lattice recommendation method based on user feedback and search content, the method uses user-input search information to build concept lattices, filters keywords, performs grouping and reduction, constructs a recommendation rating matrix, and performs predictive rating based on user reviews to improve recommendation accuracy.

Benefits of technology

It improves the recommendation accuracy of recommendation systems in IoT environments, reduces selection costs, provides higher quality recommended content, and adapts to heterogeneity and uncertainty.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116204721B_ABST
    Figure CN116204721B_ABST
Patent Text Reader

Abstract

This invention discloses a concept grid recommendation method and apparatus based on user feedback and search content. The method includes: constructing a concept grid of relevant text content based on user-input search information; constructing a formal background of implicit keywords based on user-input search content, and filtering list keywords based on the similarity between implicit keywords; constructing a concept grid based on user feedback and user characteristics, and performing grouping and reduction to construct a recommendation rating matrix for user items; predicting and rating recommended items and user comments based on the filtered list keywords, and substituting the ratings into the recommendation rating matrix to score the unrated items of the predicted target user, thereby completing the recommendation. In the current network environment, this invention can play an important role in helping users select more cost-effective, higher-quality, and more suitable recommended content, especially in the context of recommendation system management services, heterogeneity, and dimensional uncertainty.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The technical field of this invention is the Internet of Things (IoT) recommendation technology, and in particular, it relates to a concept grid recommendation method and apparatus based on user record feedback and search content. Background Technology

[0002] The amount of data on the internet and the number of internet users have been growing at an unprecedented rate. This extraordinary growth makes it crucial to improve internet users' ability to distinguish between relevant and irrelevant information.

[0003] Recommendation systems have recently been widely adopted across various fields to recommend items of interest to users. These systems help users find the most relevant information they need by requiring less time and effort.

[0004] Recommendation algorithms are susceptible to popularization bias, and the quality of recommendations can vary across demographic groups. Recommendation systems and search tools increasingly guide us to access information online, including news, entertainment, academic resources, and social connections. When evaluating the quality of these results, average performance across all users is typically reported. Therefore, while the majority of the population tends to dominate overall statistics when measuring the utility of search and recommendation tools, utility can also vary by individual and demographic group.

[0005] Recommender systems play a crucial role in providing users with useful information, especially in e-commerce applications. Many proposed models use users' rating history to predict unknown ratings. In recent years, user reviews, as a valuable source of knowledge, have attracted the attention of researchers in this field, leading to the emergence of new review-based recommender systems. Collaborative filtering, as the most widely used single-domain recommender technique, has received significant attention from both academia and industry. Matrix factorization, as one of the most popular collaborative filtering methods, has also become a hot topic. However, these techniques still face some challenges, including: data sparsity: Firstly, users rarely rate items, and due to the vast variety of items, the rating data typically used in matrix factorization suffers from sparsity, which reduces the accuracy of rating predictions in traditional collaborative filtering techniques. Secondly, the cold start problem: Collaborative filtering requires a large amount of historical interaction with the website to provide high-quality recommendations, resulting in the user cold start problem. In a new domain, users typically have little or no history of interaction with the website. Therefore, traditional recommendation methods often fail to provide high-quality recommendations for users starting from scratch. Noise problem: Noise in a recommendation system refers to data in the dataset that affects the prediction of rating values. Noise in the dataset of a recommendation system is divided into malicious noise and non-malicious noise (natural noise). Both types of noise are important and will have an adverse effect on recommendation performance.

[0006] Several mature recommendation algorithms have been developed, broadly categorized into collaborative filtering, content-based recommendation, and hybrid recommendation algorithms. Formal Concept Analysis (FCA) is an important technique for analyzing and extracting formal knowledge from a set of objects and their attributes. The extracted ontology is called a formal concept, which explains the closure relationships between objects and attributes. Due to its close connection with order theory, it can be further organized into an intuitive hierarchical structure called a concept lattice. Therefore, it is widely used in data interpretation, visualization, and analysis. Summary of the Invention

[0007] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0008] In view of the above-mentioned problems, the present invention is proposed.

[0009] Therefore, the technical problem solved by this invention is that traditional recommendation system methods fail to utilize the ever-growing, dynamic, and heterogeneous IoT data when building IoT recommendation systems.

[0010] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0011] In a first aspect, embodiments of the present invention provide a concept grid recommendation method based on user record feedback and search content, including:

[0012] Construct a conceptual grid of relevant text content based on the search information entered by the user;

[0013] The form background of implicit keywords is constructed based on the search content input by the user, and the list of keywords is filtered according to the similarity between the implicit keywords;

[0014] Based on user feedback records and user characteristics, a concept lattice is constructed, and then grouped and reduced to build a recommendation rating matrix for user items;

[0015] Based on the selected list keywords, the recommended items and user reviews are predicted and rated. The ratings are then substituted into the recommendation rating matrix to score the unrated items for the target users, thus completing the recommendation.

[0016] As a preferred embodiment of the concept lattice recommendation method based on user record feedback and search content, the construction of the concept lattice of relevant text content includes: constructing a formal background T(U,G,I) of implicit keywords based on the search content input by the user, where U is the set of users, G is the set of keywords, and I is the binary relation between U and G.

[0017] Let the association matrix between the category label set B and the keyword set G be T = {t} ij}, when B i Contains keyword G j At that time, t ij =1, otherwise t ij =0, T is transposed to obtain the association matrix T between the keyword set G and the category label set B. B ;

[0018] Based on the association matrix T between the keyword set G and the category label set B B Obtain a lexical representation based on category labels:

[0019]

[0020] Where, ν ij The value represents the keyword g. i Is it included in category label B? j In the middle, when B j does not contain g i At that time, ν ij =0, otherwise ν ij =1, which is represented as:

[0021] g i ={B j |g i ∈B j}

[0022] Let keywords g1 and g2 be given, where g1 corresponds to m category labels, denoted as g1 = {B}. 11 B 12 ,...,B 1m}, g2 corresponds to n category labels, represented as g2={B 21 B 22 ,...,B 2n}

[0023] As a preferred embodiment of the concept grid recommendation method based on user record feedback and search content, the filtering list keywords include: calculating the category label B2 with the highest similarity to the category label B1 of the keyword input by the user, and calculating the similarity between other keywords belonging to B2 and the keyword input by the user. If the similarity is greater than or equal to the threshold β, the word compared with the keyword input by the user is retained; otherwise, the word is deleted from the list.

[0024] The formula for calculating the similarity between keywords can be expressed as:

[0025]

[0026] Among them, B 1i Let B be the i-th category label corresponding to g1. 2j For the j-th category label corresponding to g2, sim(B 1i B 2j ) represents category label pairs (B 1i B 2j The similarity between )

[0027] Category label B i With B j The similarity between them can be expressed as:

[0028]

[0029] Among them, B i With B j These represent nodes for two category labels in a concept lattice structure; LCS represents B. i With B j The nearest common parent node; dep(LCS) represents the depth of LCS, dis(B i B j ) indicates the path length.

[0030] As a preferred embodiment of a concept grid recommendation method based on user feedback and search content, the user feedback includes user search status and user historical behavior; the user search status includes: number of searches, number of selections, number of favorites, and number of views; the user historical behavior includes user historical item comments and ratings; and the user characteristics include search status values.

[0031] Users are grouped and categorized based on their behavior on the search recommendation page. If a user's search frequency exceeds that of 70% of users, they are defined as a frequent user; if it exceeds that of 40% of users, they are defined as a moderate user; otherwise, they are defined as a rare user.

[0032] If a user interacts with the recommended page content in two or more ways, the user is considered satisfied with the recommended content, and the score m is 5 points; if a user interacts with the recommended page content in one way, the user is considered to be moderately satisfied with the recommended content, and the score is [2,4]; if a user does not interact with the recommended page content, the user is considered to be dissatisfied with the recommended content, and the score m<2; interactive behaviors include: clicking; dwell time greater than 30 seconds; and page content collection, selection, and copying operations.

[0033] As a preferred embodiment of a concept grid recommendation method based on user record feedback and search content, the grouping reduction includes: using A T (g) represents the attribute that object g has in the formal context T, using OT (m) represents an object that has attribute m in the formal context T;

[0034] If (g, m′) ∈ T, then m′ ∈ A T (g), if (g′,m)∈T, then g′∈O T (m), for G, M represents a collection of objects / properties, A T (G) is defined as A T (g i ), O T (M) is defined as O T (m i );

[0035] When there is no ambiguity, the subscript T is omitted; additionally, x is used. ij Using double subscripts to represent binary relations (g) i ,m j The occurrence rate in the corresponding formal context; let the formal context be (G, M, I), where G has n objects and M has m attributes. In a specific grouping-based reduction, the object set G is divided into k disjoint groups S. (1) S (2) S (3) , ...,S (k) S (1) ∪S (2) ∪S (3) ∪...∪S (k) =G; classifying objects from the same group as similar, and assigning all g∈S (i) After restoration, use an object r i Replacement, defining the reduced background fidelity as:

[0036]

[0037] The integer linear programming technique is used to solve the concept lattice reduction based on grouping. The integer linear programming model is defined as follows:

[0038] Maximize c T x

[0039] subject to Ax≤b,x≥0 and x∈Z n

[0040] Here, x is called a variable, which is the vector to be determined, Z represents an integer, and A, b, and c are coefficient matrices or vectors that constitute constraints.

[0041] The global objective of this integer linear programming model is to maximize T. The variable allocation rule is as follows: for object g...i rate of change x ij =B, if there is another object g k ,satisfy:

[0042] |A(g i )-A(g k ))Y(A(g k )-A(g i ))|≤ε r ;

[0043] Where ε r The threshold value is set; and x kj =1-B, then x ij Marked as mutable, it is called c. ij And assign a variable v ij To represent x ij Whether it has been modified; where B is set to true for the split rule and false for the join rule;

[0044] The basic constraints of this integer linear programming model are:

[0045]

[0046]

[0047] In addition, the following constraints need to be added to maintain formal background fidelity:

[0048]

[0049] Where ε m The threshold value is set.

[0050] As a preferred embodiment of the concept lattice recommendation method based on user record feedback and search content, the construction of the user item recommendation rating matrix includes: converting the recommendation rating matrix into a binary rating matrix, defining a rating of 4-5 as recommended and setting it to 1, and defining the remaining ratings as not recommended and setting it to 0. This matrix is ​​regarded as the formal background and represented as C(U,I,S), where U, I, and S represent the user set, the item set, and the relationship between users and items, respectively. After obtaining the user interest formal background, a concept lattice structure model is established based on the binary relationship between objects and attributes in the user interest formal background C.

[0051] As a preferred embodiment of a concept grid recommendation method based on user feedback and search content, the predicted rating includes: denoting the membership degree of a given review r in the i-th item as μ based on the rating. i (r), and computed using the following simple heuristic:

[0052]

[0053] Wherein, the actual ratio (r) represents the actual rating score corresponding to review r in the dataset;

[0054] Let f represent the k-th fuzzy set belonging to a user u and a project i. k (u) and f k (i), 1≤k≤5, convert each comment into a semantic vector, each f k (u) and f k (i) There is a set of vectors for f k The average of the included vectors is taken to obtain a single vector, resulting in 5 vectors for user u and 5 vectors for item i. Each vector corresponds to one of the possible rating values, denoted as V. k (u) and V k (i); for V k (u) and V k (i) The calculation is as follows:

[0055] V k (u)=Avg(μ k (r,u)×V k (r,u))

[0056] V k (i)=Avg(μ k (r,i)×V k (r,i))

[0057] Among them, V k (r,u) represents user u's relationship to f. k The semantic vector of the r-th comment of (u), μ k (r,u) is f k Membership degree in (u);

[0058] By comparing the vectors of user u and item i for each corresponding rating value, the rating value with the highest similarity is considered the predicted rating for the pair of u and i:

[0059]

[0060] Among them, V k (u) and V k (i) are vectors belonging to user u and the i-th item, respectively, corresponding to the rating score of k, and r is the predicted rating of u for i.

[0061] Secondly, embodiments of the present invention provide a concept grid recommendation system based on user record feedback and search content, characterized in that it includes:

[0062] The text content concept grid construction module is used to construct concept grids of relevant text content based on the search information input by the user.

[0063] The filtering module is used to construct a background of implicit keywords based on the user's search input, and to filter the list of keywords based on the similarity between the implicit keywords;

[0064] The user concept lattice construction module is used to construct concept lattices based on user feedback records and user characteristics, and to perform grouping and reduction to build a recommendation rating matrix for user items;

[0065] The prediction and recommendation module is used to predict and rate recommended items and user reviews based on the selected list keywords. The ratings are then substituted into the recommendation rating matrix to score the unrated items for the target user, thereby completing the recommendation.

[0066] Thirdly, embodiments of the present invention provide a computing device, including:

[0067] Memory and processor;

[0068] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the one or more programs are executed by the one or more processors, the one or more processors implement the concept grid recommendation method based on user record feedback and search content as described in any embodiment of the present invention.

[0069] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the aforementioned concept grid recommendation method based on user record feedback and search content.

[0070] The beneficial effects of this invention are as follows: This invention proposes a concept lattice recommendation method based on user record feedback and search content. By constructing a concept lattice from user record feedback and search content, user objects are grouped and the concept lattice is reduced, thereby improving recommendation accuracy. In the current network environment, given the uncertainty in recommendation system management services, heterogeneity, and dimensions, it can play an important role in helping users select more cost-effective, higher-quality, and more suitable recommended content. Attached Figure Description

[0071] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0072] Figure 1 This is an overall flowchart of the concept grid recommendation method based on user record feedback and search content described in the first embodiment of the present invention;

[0073] Figure 2 This is a schematic diagram comparing the MAE of different methods in the simulation and control experiment of the concept grid recommendation method based on user record feedback and search content described in the second embodiment of the present invention;

[0074] Figure 3 This is a schematic diagram of the results based on CC metric in the simulation and control experiment of the concept grid recommendation method based on user record feedback and search content described in the second embodiment of the present invention. Detailed Implementation

[0075] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0076] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0077] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0078] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0079] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0080] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0081] Example 1

[0082] Reference Figure 1 This is the first embodiment of the present invention, which provides a concept grid recommendation method based on user record feedback and search content, including:

[0083] S1: Construct a conceptual grid of relevant text content based on the search information entered by the user;

[0084] Specifically, the concept lattice for constructing relevant text content includes: constructing a formal background T(U,G,I) of implicit keywords based on the search content input by the user, where U is the set of users, G is the set of keywords, and I is a binary relation between U and G;

[0085] Let the association matrix between the category label set B and the keyword set G be T = {t} ij}, when B i Contains keyword G j At that time, t ij =1, otherwise t ij =0, T is transposed to obtain the association matrix T between the keyword set G and the category label set B. B ;

[0086] Based on the association matrix T between the keyword set G and the category label set B B Obtain a lexical representation based on category labels:

[0087]

[0088] Where, ν ij The value represents the keyword g. i Is it included in category label B? jIn the middle, when B j does not contain g i At that time, ν ij =0, otherwise ν ij =1, which is represented as:

[0089] g i ={B j |g i ∈B j}

[0090] Let keywords g1 and g2 be given, where g1 corresponds to m category labels, denoted as g1 = {B}. 11 B 12 ,...,B 1m}, g2 corresponds to n category labels, represented as g2={B 21 B 22 ,...,B 2n}

[0091] It should be noted that, because recommender systems require sufficient information to generate accurate recommendations, their performance in predicting unseen items largely depends on the keywords used in each user's search. Generally, the accuracy of recommendations may decrease if users' search keywords are too simple. In this step, we propose an effective strategy to enrich the content searched by users.

[0092] When a user enters a query, relevant recommendations must be retrieved from a vast pool of available information in the registry, a task that remains challenging. Consequently, extensive research is underway in the field of recommender systems.

[0093] Formal concept analysis can be viewed as analyzing the numerical structure of data through concepts. A group of objects reflecting a set of common attributes is called a concept. The group of objects with similar attributes in a concept is called the extension of the concept, while the attributes that construct the hierarchical structure of the concept are called the intension.

[0094] It should also be noted that the relevance of user-input text can be measured through category tags. When keywords appear together in a single tag, it indicates a certain degree of correlation. When these keywords appear simultaneously in multiple category tags, it indicates a strong correlation. The similarity between keywords g1 and g2 is influenced by the number of category tags for that keyword and the semantic relevance between different tags within those categories. Keyword similarity is directly proportional to the number of category tags and the degree of correlation between tags.

[0095] S2: Construct a background of implicit keywords based on the user's search input, and filter the list of keywords based on the similarity between the implicit keywords;

[0096] Furthermore, the keyword filtering list includes: calculating the category label B2 that has the highest similarity to the category label B1 of the keyword input by the user, and calculating the similarity between other keywords belonging to B2 and the keyword input by the user. If the similarity is greater than or equal to the threshold β, the word compared with the keyword input by the user is retained; otherwise, the word is deleted from the list.

[0097] The formula for calculating the similarity between keywords can be expressed as:

[0098]

[0099] Among them, B 1i Let B be the i-th category label corresponding to g1. 2j For the j-th category label corresponding to g2, sim(B 1i B 2j ) represents category label pairs (B 1i B 2j The similarity between )

[0100] Category label B i With B j The similarity between them can be expressed as:

[0101]

[0102] Among them, B i With B j These represent nodes for two category labels in a concept lattice structure; LCS represents B. i With B j The nearest common parent node; dep(LCS) represents the depth of LCS, dis(B i B j ) indicates the path length.

[0103] S3: Construct a concept lattice based on user feedback records and user characteristics, and perform grouping and reduction to build a recommendation rating matrix for user items;

[0104] Specifically, the user record feedback includes user search status and user historical behavior; the user search status includes: number of searches, number of selections, number of favorites, and number of views; the user historical behavior includes user's historical project comments and ratings; the user characteristics include search status values;

[0105] Users are grouped and categorized based on their behavior on the search recommendation page. If a user's search frequency exceeds that of 70% of users, they are defined as a frequent user; if it exceeds that of 40% of users, they are defined as a moderate user; otherwise, they are defined as a rare user.

[0106] If a user interacts with the recommended page content in two or more ways, the user is considered satisfied with the recommended content, and the score m is 5 points; if a user interacts with the recommended page content in one way, the user is considered to be moderately satisfied with the recommended content, and the score is [2,4]; if a user does not interact with the recommended page content, the user is considered to be dissatisfied with the recommended content, and the score m<2; interactive behaviors include: clicking; dwell time greater than 30 seconds; and page content collection, selection, and copying operations.

[0107] It should be noted that reviews provided by different users for a specific project can represent the key characteristics of that project; the history of user-written reviews can indicate user preferences. This invention argues that measuring the similarity between user-written reviews and reviews provided for a project can reveal the similarity between user preferences and project content.

[0108] It should also be noted that the present invention proposes a reliable method for analyzing group-based conceptual lattice reduction, mainly considering two aspects: formal background fidelity and lattice fidelity.

[0109] Specifically, the grouping reduction includes: using A T (g) represents the attribute that object g has in the formal context T, using O T (m) represents an object that has attribute m in the formal context T;

[0110] If (g, m′) ∈ T, then m′ ∈ A T (g), if (g′,m)∈T, then g′∈O T (m), for G, M represents a collection of objects / properties, A T (G) is defined as A T (g i ), O T (M) is defined as O T (m i );

[0111] When there is no ambiguity, the subscript T is omitted; additionally, x is used. ij Using double subscripts to represent binary relations (g) i ,m j The occurrence rate in the corresponding formal context; let the formal context be (G, M, I), where G has n objects and M has m attributes. In a specific grouping-based reduction, the object set G is divided into k disjoint groups S. (1) S (2) S (3) , ...,S (k) S (1) ∪S(2) ∪S (3) ∪...∪S (k) =G; classifying objects from the same group as similar, and assigning all g∈S (i) After restoration, replace it with an object ri, and define the reduced background fidelity as:

[0112]

[0113] The integer linear programming technique is used to solve the concept lattice reduction based on grouping. The integer linear programming model is defined as follows:

[0114] Maximize c T x

[0115] subject to Ax≤b,x≥0 and x∈Z n

[0116] Here, x is called a variable, which is the vector to be determined, Z represents an integer, and A, b, and c are coefficient matrices or vectors that constitute constraints.

[0117] It should be noted that the main objective of the alternating form of group-based reduction is to maximize the increase of those with the same opponent. i The number of objects o, i.e., A(o) = A(o) i This objective can be represented by the following global objective function. Here, I ij It is a variable indicating whether i and j are the same after modification, and n is the size of the original object set.

[0118]

[0119] This function calculates elements that have at least one common element o. j The number of objects, where j>i. It can be easily derived that this is equal to the number of objects reduced from the original form based on grouping. Therefore, the global objective of the ILP model is to maximize T.

[0120] In theory, all variable elements in a formal context can be modified, so we can assign a total of n × m variables to a formal context with n objects and m properties. However, we tend to modify only objects that are already similar. Excessive modification of the formal context, however, degrades its fidelity and incurs significant computational costs.

[0121] This preference can be formally described as, only when |A T We only consider modifying objects g1 and g2 to make A(g1) = A(g2) when |(g1) - A(g2)| is less than the threshold ε. That is, for an object g1, if for any other object g2 in the original formal context, we have |A|...T (g1)-A(g2)|>ε, the object is marked as immutable, and we will not assign any variables to the events representing the g1 property.

[0122] Specifically, the global objective of this integer linear programming model is to maximize T, and the variable allocation rule is as follows: for object g i rate of change x ij =B, if there is another object g k ,satisfy:

[0123] |(A(g i )-A(g k ))Y(A(g k )-A(g i ))|≤ε r ;

[0124] Where ε r The threshold value is set; and x kj =1-B, then x ij Marked as mutable, it is called c. ij And assign a variable v ij To represent x ij Whether it has been modified; where B is set to true for the split rule and false for the join rule;

[0125] The basic constraints of this integer linear programming model are:

[0126]

[0127]

[0128] In addition, the following constraints need to be added to maintain formal background fidelity:

[0129]

[0130] Where ε m The threshold value is set.

[0131] It should be noted that, to ensure the fidelity of the concept lattice, this method strictly follows the grouping-merging process, employs a separation or connection strategy for merging, and applies the merging strategy globally.

[0132] For any formal background T = (G, M, I), if we have g1, g2 ∈ G and A(g1) = A(g2), then removing g1 or g2 from G will not change the conceptual lattice structure. That is, T′ = (G - {g1}, M, I), T″ = (G - {g2}, M, I), thus L(T), L(T′), and L(T″) have the same structure.

[0133] The construction of the user item recommendation rating matrix includes: converting the recommendation rating matrix into a binary rating matrix, defining a rating of 4-5 as recommended (set to 1), and defining the remaining ratings as not recommended (set to 0). This matrix is ​​regarded as a formal background, represented as C(U,I,S), where U, I, and S represent the user set, the item set, and the relationship between users and items, respectively. After obtaining the user interest formal background, a conceptual lattice structure model is established based on the binary relationship between objects and attributes in the user interest formal background C.

[0134] S4: Based on the recommended items and user reviews obtained from the selected list keywords, predict the ratings and substitute the ratings into the recommendation rating matrix to score the unrated items of the target users, thereby completing the recommendation.

[0135] Specifically, the predicted rating includes: denoting the membership degree of a given review r in the i-th item as μ based on the rating. i (r), and computed using the following simple heuristic:

[0136]

[0137] Wherein, the actual ratio (r) represents the actual rating score corresponding to review r in the dataset;

[0138] It should be noted that, due to the uncertainty in assigning scores, a fuzzy segmentation process is used to divide the entire review process into five groups based on the scores. In this method, a reviewer can belong to multiple sets, but with different degrees of membership.

[0139] It should also be noted that in this method, we utilize the Word2vec embedding model to construct a semantic vector for each corpus. Word2vec is similar to an autoencoder, encoding each word into a vector, but instead of training on the input word through reconstruction, it is trained against other neighboring words in the corpus. In this model, using a training corpus, a target word and each word in its context are transformed into a popular vector of vocabulary size or a deterministic subset thereof. In this vector, there is only a "1" at the position corresponding to the target word, and only 0s at all other positions. These vector sets serve as the input layer of the neural network. The output of the network is a single vector for each word, containing the probability that a randomly selected neighboring word is that word in the corpus.

[0140] After creating a vector for each word, an aggregate vector representing the entire review text should be constructed. There are various methods to obtain the final vector of a document, including vector summation, averaging, and mean averaging. This method uses the mean averaging of the vectors to obtain the final vector.

[0141] Since only two single vectors representing users and projects are needed respectively, the vectors contained in a set should be aggregated into a single vector for this purpose. Because the membership degrees of vectors are generally unequal within a set, as a first step, each vector needs to be multiplied by its membership degree. Therefore, lower membership degrees have less impact on constructing the final vector. Then, for f... k The average of the vectors contained in the vector is taken to obtain a single vector.

[0142] Specifically, the k-th fuzzy set belonging to a user u and an item i is represented as f. k (u) and f k (i), 1≤k≤5, convert each comment into a semantic vector, each f k (u) and f k (i) There is a set of vectors for f k The average of the included vectors is taken to obtain a single vector, resulting in 5 vectors for user u and 5 vectors for item i. Each vector corresponds to one of the possible rating values, denoted as V. k (u) and V k (i); for V k (u) and V k (i) The calculation is as follows:

[0143] V k (u)=Avg(μ k (r,u)×V k (r,u))

[0144] V k (i)=Avg(μ k (r,i)×V k (r,i))

[0145] Among them, V k (r,u) represents user u's relationship to f. k The semantic vector of the r-th comment of (u), μ k (r,u) is f k Membership degree in (u);

[0146] By comparing the vectors of user u and item i for each corresponding rating value, the rating value with the highest similarity is considered the predicted rating for the pair of u and i:

[0147]

[0148] Among them, V k (u) and V k (i) are vectors belonging to user u and the i-th item, respectively, corresponding to the rating score of k, and r is the predicted rating of u for i.

[0149] Example 2

[0150] Reference Figure 2-3 As an embodiment of the present invention, a concept grid recommendation method based on user record feedback and search content is provided. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation and control experiments.

[0151] The dataset used in this study is the Amazon dataset (https: / / www.amazon.com / movies-tv-dvd-bluray / b / ref=sd_allcat_mov?ie=UTF8&node=2625373011). Mean absolute error (MAE), catalog coverage (CC), diversity, and novelty were used as four well-known evaluation metrics.

[0152] Table 1 reports the experimental results using the movie dataset, where our method achieves the best MAE, CC, diversity, and novelty values ​​compared to other methods. Performance comparisons of recommendation methods based on the Amazon dataset are shown, with the best results in bold and the second-best results underlined. Clearly, our method can significantly improve the quality of recommendations across various evaluation metrics.

[0153] Table 1 shows the experimental results using the movie dataset.

[0154]

[0155] To verify the effectiveness of the proposed method in sparse scenarios, the available ratings in the sample dataset were randomly changed to zero, creating five datasets with different sparsities: 93%, 96%, 97%, 98%, and 99%. The proposed method and existing PCC re-prediction methods were then tested on these five datasets with varying sparsities. Figure 2 MAE results from different methods were compared.

[0156] The number of items recommended to users in the recommendation list (N in the top-N) affects the obtained values ​​of diversity, novelty, and CC (Conversion Control) metrics. Therefore, experiments were conducted to investigate the impact of different values ​​of the parameter N on the recommender's performance. Figure 3 This is the result based on the CC metric. The graph shows that CC increases as n increases.

[0157] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A concept grid recommendation method based on user record feedback and search content, characterized in that, include: Construct a conceptual grid of relevant text content based on the search information entered by the user; The concept lattice for constructing relevant text content includes: constructing a formal background T(U,G,I) of implicit keywords based on the search content input by the user, where U is the set of users, G is the set of keywords, and I is a binary relation between U and G; Let the association matrix between the category label set B and the keyword set G be... ,when Contains keywords hour, ,otherwise Transpose T to obtain the association matrix between the keyword set G and the category label set B. ; Based on the association matrix between the keyword set G and the category label set B Obtain a lexical representation based on category labels: in, Value represents keyword Is it included in the category label? In the middle, when Not included hour, ,otherwise , is represented as: Set keywords and ,in For m category labels, it is represented as , For n category labels, it is represented as ; The form background of implicit keywords is constructed based on the search content input by the user, and the list of keywords is filtered according to the similarity between the implicit keywords; Based on user feedback records and user characteristics, a concept lattice is constructed, and then grouped and reduced to build a recommendation rating matrix for user items; The grouping reduction includes: using This represents the properties that object g has in the formal context T, using... This represents an object with property m in a formal context T; For G, M represents a collection of objects / properties. Defined as , Defined as ; When there is no ambiguity, the subscript T is omitted; additionally, use... Using double subscripts to represent binary relations The occurrence rate in the corresponding formal context; let the formal context be (G, M, I), where G has n objects and M has m attributes. In a specific grouping-based reduction, the set of objects G is divided into k disjoint groups. ,in Objects from the same group are considered similar; all of them are considered similar. After restoration, use an object Replacement, defining the reduced background fidelity as: The integer linear programming model is defined as follows: This method utilizes integer linear programming techniques to solve the concept lattice reduction based on grouping. Here, x is called a variable, which is the vector to be determined, Z represents an integer, and A, b, and c are coefficient matrices or vectors that constitute constraints. The global objective of this integer linear programming model is to maximize T. The variable allocation rule is as follows: for object g... i rate of change If there is another object g k, satisfy: ; in ε r The threshold value is set; and Then x ij Marked as mutable, it is called c. ij And assign a variable v ij To represent x ij Whether it has been modified; where B is set to true for the split rule and false for the join rule; The basic constraints of this integer linear programming model are: In addition, the following constraints need to be added to maintain formal background fidelity: ; in ε m The set threshold; Based on the selected list keywords, the recommended items and user reviews are predicted and rated. The ratings are then substituted into the recommendation rating matrix to score the unrated items for the target users, thus completing the recommendation.

2. The concept grid recommendation method based on user record feedback and search content as described in claim 1, characterized in that, The keywords in the filtering list include: calculating the category label B2 that has the highest similarity to the category label B1 of the keyword entered by the user, and calculating the similarity between other keywords belonging to B2 and the keyword entered by the user. If the similarity is greater than or equal to the threshold β, the word compared with the keyword entered by the user is retained; otherwise, the word is deleted from the list. The formula for calculating the similarity between keywords is as follows: in, for The corresponding i-th category label, for The corresponding j-th category label, Indicates category label pairs Similarity; Category Tags and The similarity between them is expressed as: in, and These represent nodes representing two category labels in a concept lattice structure; LCS represents... and The nearest common parent node; Indicates the depth of the LCS. Indicates the path length.

3. The concept grid recommendation method based on user record feedback and search content as described in claim 2, characterized in that, The user record feedback includes user search status and user historical behavior; the user search status includes: number of searches, number of selections, number of favorites, and number of views; the user historical behavior includes user's historical project comments and ratings; the user characteristics include search status values; Users are grouped and categorized based on their behavior on the search recommendation page. If a user's search frequency exceeds 70% of users, they are defined as a frequent user; if it exceeds 40% of users, they are defined as a moderate user; otherwise, they are defined as a rare user. If a user interacts with the recommended page content in two or more ways, the user is considered satisfied with the recommended content, and the score m is 5 points; if a user interacts with the recommended page content in one way, the user is considered to be moderately satisfied with the recommended content, and the score is [2,4]; if a user does not interact with the recommended page content, the user is considered to be dissatisfied with the recommended content, and the score m<2; interactive behaviors include: clicking; dwell time greater than 30 seconds; and page content collection, selection, and copying operations.

4. The concept grid recommendation method based on user record feedback and search content as described in claim 3, characterized in that, The construction of the user item recommendation rating matrix includes: converting the recommendation rating matrix into a binary rating matrix, defining a rating of 4-5 as recommended and setting it to 1, and defining the remaining ratings as not recommended and setting them to 0. This matrix is ​​regarded as a formal background, represented as C(U, I, S), where U, I, and S represent the user set, the item set, and the relationship between users and items, respectively. After obtaining the user interest formal background, a conceptual lattice structure model is established based on the binary relationship between objects and attributes in the user interest formal background C.

5. The concept grid recommendation method based on user record feedback and search content as described in claim 4, characterized in that, The predicted rating includes: recording the membership degree of a given review r in the i-th project based on the rating. And it is calculated using the following simple heuristic: Wherein, the actual ratio (r) represents the actual rating score corresponding to r in the dataset; Let the k-th fuzzy set belonging to a user u and an item i be represented as... and 1≤k≤5, convert each comment into a semantic vector, each and There is a set of vectors, for The average of the included vectors is taken to obtain a single vector, resulting in 5 vectors for user u and 5 vectors for item i. Each vector corresponds to one of the possible rating values, denoted as . and ;right and The calculation is as follows: in, It belongs to user u for The semantic vector of the r-th comment, yes Membership degree in; By comparing the vectors of user u and item i for each corresponding rating value, the rating value with the highest similarity is considered the predicted rating for the pair of u and i: in, and These are vectors belonging to user u and the i-th item, respectively, corresponding to the rating score of k. It is the predicted score of u for i.

6. A concept grid recommendation system based on user record feedback and search content, characterized in that, include: The text content concept grid construction module is used to construct concept grids of relevant text content based on the search information input by the user. The concept lattice for constructing relevant text content includes: constructing a formal background T(U,G,I) of implicit keywords based on the search content input by the user, where U is the set of users, G is the set of keywords, and I is a binary relation between U and G; Let the association matrix between the category label set B and the keyword set G be... ,when Contains keywords hour, ,otherwise Transpose T to obtain the association matrix between the keyword set G and the category label set B. ; Based on the association matrix between the keyword set G and the category label set B Obtain a lexical representation based on category labels: in, Value represents keyword Is it included in the category label? In the middle, when Not included hour, ,otherwise , is represented as: Set keywords and ,in For m category labels, it is represented as , For n category labels, it is represented as ; The filtering module is used to construct a background of implicit keywords based on the user's search input, and to filter the list of keywords based on the similarity between the implicit keywords; The user concept lattice construction module is used to construct concept lattices based on user feedback records and user characteristics, and to perform grouping and reduction to build a recommendation rating matrix for user items; The grouping reduction includes: using This represents the properties that object g has in the formal context T, using... This represents an object with property m in a formal context T; For G, M represents a collection of objects / properties. Defined as , Defined as ; When there is no ambiguity, the subscript T is omitted; additionally, use... Using double subscripts to represent binary relations The occurrence rate in the corresponding formal context; let the formal context be (G, M, I), where G has n objects and M has m attributes. In a specific grouping-based reduction, the set of objects G is divided into k disjoint groups. ,in Objects from the same group are considered similar; all of them are considered similar. After restoration, use an object Replacement, defining the reduced background fidelity as: The integer linear programming model is defined as follows: This method utilizes integer linear programming techniques to solve the concept lattice reduction based on grouping. Here, x is called a variable, which is the vector to be determined, Z represents an integer, and A, b, and c are coefficient matrices or vectors that constitute constraints. The global objective of this integer linear programming model is to maximize T. The variable allocation rule is as follows: for object g... i rate of change If there is another object g k, satisfy: ; in ε r The threshold value is set; and Then x ij Marked as mutable, it is called c. ij And assign a variable v ij To represent x ij Whether it has been modified; where B is set to true for the split rule and false for the join rule; The basic constraints of this integer linear programming model are: In addition, the following constraints need to be added to maintain formal background fidelity: ; in ε m The set threshold; The prediction and recommendation module is used to predict and rate recommended items and user reviews based on the selected list keywords. The ratings are then substituted into the recommendation rating matrix to score the unrated items for the target user, thereby completing the recommendation.

7. A computing device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the concept grid recommendation method based on user record feedback and search content as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the concept grid recommendation method based on user record feedback and search content as described in any one of claims 1 to 5.