A data retrieval method based on multi-element matrix fusion

Abstract

The application relates to the technical field of data retrieval, in particular to a data retrieval method based on multi-element matrix fusion, which comprises the following steps: S1, multi-element classification, a whole set of source data is split into a plurality of independent data elements according to a unified classification standard through defining classification dimensions; the classification granularity, that is, the fineness degree of classification, is controlled in the classification process, and a consistency check is performed to ensure that the granularity is reasonable, thereby establishing a two-dimensional matrix construction basis for each element; S2, matrix construction, based on the independent data elements obtained through the multi-element classification in S1; and S3, similar interval determination, based on the multi-type matrix constructed in S2; in the scheme, the data elements are split according to a unified standard, the multi-type matrix is constructed, the standardized instruction analysis and similar interval determination are performed, the weighted fusion calculation of the similar level and the optimization result presentation are performed, the structured processing of multi-dimensional data, the accurate collaborative retrieval and the decision auxiliary support are realized, and the problems of lack of systematic classification and multi-dimensional collaboration in the data processing of the retrieval method are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data retrieval, in particular to a data retrieval method based on multi-element matrix fusion. BACKGROUND

[0002] E-commerce data retrieval is the core link of e-commerce platform to improve user experience and transaction conversion. With the development of e-commerce business, the platform has accumulated massive multi-type data, including user click, collection, stay behavior data, commodity price, sales, text description attribute data, and transaction amount, after-sales status record data, and user retrieval demand has changed from single numerical filtering to "numerical + semantic + behavior" multi-dimensional composite demand, which puts forward higher requirements for retrieval accuracy and decision assistance.

[0003] The traditional data retrieval method has the problems of lack of systematic classification of data processing and insufficient multi-dimensional coordination. Different types of data are often mixed and integrated, and the data attributes or business modules are not split into independent data elements and a unified type matrix is not constructed, resulting in confusion of similar determination standards and easy retrieval deviation. In addition, the output result depends on a single dimension, or the multi-dimensional data is simply spliced, without combining business requirements to allocate weights and unify similar levels, which cannot comprehensively reflect the user's multi-faceted needs, and the result matching is poor.

[0004] Therefore, the present application provides a data retrieval method based on multi-element matrix fusion, which splits data elements according to a unified standard and constructs multi-type matrix, standardizes instruction analysis and similar interval determination, weighted fusion calculation of similar level and optimization of result presentation, realizes the structured processing of multi-dimensional data, accurate collaborative retrieval and decision assistance support, and solves the problems of lack of systematic classification of data processing and insufficient multi-dimensional coordination of traditional retrieval methods. SUMMARY

[0005] The technical problem solved is the lack of systematic classification of data processing and insufficient multi-dimensional coordination.

[0006] In view of the deficiencies in the prior art, the present application provides a data retrieval method based on multi-element matrix fusion, thereby solving the technical problems mentioned in the background art.

[0007] To achieve the above purpose, the present application realizes the following technical scheme:

[0008] A data retrieval method based on multi-element matrix fusion, comprising the following steps:

[0009] S1, multi-element classification, the whole set of source data is split into multiple independent data elements according to a unified classification standard by defining classification dimensions, the classification process controls the classification granularity, that is, the fineness of the classification, and the consistency check ensures that the granularity is reasonable, and a two-dimensional matrix is established for the subsequent elements to build the foundation;

[0010] S2, matrix construction, based on the independent data elements obtained by S1 multivariate classification, a two-dimensional matrix (X, Y) is built for each element, which is unified in structure and can be directly used for retrieval matching;

[0011] S3, similar interval determination, based on the multi-type matrix constructed in S2, through instruction conversion condition, matching corresponding matrix, positioning target coordinates, unified determination standard, the similar data meeting the user's retrieval requirements is filtered out from the matrix, providing accurate input for subsequent multi-matrix data fusion;

[0012] S4, fusion calculation layer, based on the multi-matrix data aligned by S3 similar interval determination, by assigning weights to each aligned matrix data, the final fusion value is calculated, the multi-dimensional matching degree is divided according to the fusion value, the similarity interval is divided, and the optimal choice is quickly locked for the retrieval results of the similarity interval division;

[0013] S5, retrieval presentation, based on the whole process of S1 to S4, the final output result list can be used for direct decision.

[0014] In one possible implementation, in S1:

[0015] The classification dimension includes two types: classification according to the attributes / types of data, which divides the data of different attributes in the source data, and the source data includes user click times, product prices, product description text, and user collection status, which are divided into user click behavior data, product numerical attribute data, product text description data, and user collection behavior data; classification according to business scenarios / retrieval purposes, combined with specific retrieval scenarios, in the e-commerce product matching scenario, the source data is divided into user behavior data, product attribute data, and transaction record data; each data in the source data belongs to only one data element, without omission or duplication;

[0016] The specific process of classification granularity is: first, coarse division according to the core dimension, based on the selected classification dimension, the source data is divided into 2-3 coarse granularity elements, in the e-commerce scenario, into user behavior data, product attribute data, and evaluation data; then, it is judged whether each coarse granularity element meets the data type uniformity and can be directly used for S2 to construct a two-dimensional matrix, when the data types are mixed, the coarse granularity element is finely divided until the data type and purpose of each element are uniform; the total number of S2 two-dimensional matrices corresponding to the finally determined medium granularity data element is ≤5, and at most not more than 8, avoiding the dramatic increase in calculation amount due to too many matrices in subsequent S4 fusion calculation.

[0017] In one possible implementation, the axis definition of the two-dimensional matrix (X, Y) in S2 needs to follow "unique identification to determine the subject, and association object / feature to determine the dimension":

[0018] The X axis or the Y axis contains a unique identifier such as a user ID, a product ID, etc., to ensure that the "subject-association object" combination is unique, the axis dimension is stable, and covers all core subjects / features of the corresponding data element;

[0019] The matrix is divided into numerical type, Boolean type and text encoding type according to the type of value, and each matrix belongs to only one type.

[0020] In a possible implementation, S2 further includes a matrix data filling step:

[0021] In the missing value processing, the numerical type matrix is filled with 0 first, and the missing key attribute is filled with the median of the same type data; the Boolean type matrix is uniformly filled with 0; the text encoding matrix is filled with a full 0 vector;

[0022] In the abnormal value processing, the numerical type matrix exceeding 3 times the standard deviation or the logically contradictory data is replaced by the 3 times standard deviation boundary value of the corresponding column; when the non-standard value appears in the Boolean type matrix or the text encoding matrix, 0 or a full 0 vector is filled.

[0023] In a possible implementation, S3 specifically includes:

[0024] The core keywords of the split user retrieval instruction are retrieved, the core keywords include a retrieval subject, a core demand and a limiting condition; the similar judgment mode is determined according to the core demand corresponding to the matrix type of S2, the numerical type demand corresponds to the threshold interval judgment, the non-class demand corresponds to the 0 / 1 matching judgment, and the semantic type demand corresponds to the similarity threshold judgment; the analysis result is converted into a standardized format of {retrieval subject}+{dimension}+{operator}+{target value / threshold value}; the industry general threshold or the user-defined supplement is adopted for the fuzzy instruction, so that the instruction is unambiguous.

[0025] In a possible implementation, S3 specifically includes:

[0026] The dimension type based on the standardized condition is matched with the two-dimensional matrix corresponding to S2, the numerical type condition is mapped to the numerical type matrix, the Boolean type condition is mapped to the Boolean type matrix, and the text semantic type condition is mapped to the text encoding matrix; when multiple conditions are combined, the corresponding matrices are respectively mapped to form a matrix combination list; all the mapped matrices need to contain the same unique identifier, so as to ensure that the subsequent data can be associated.

[0027] In a possible implementation, S3 specifically includes:

[0028] The numerical type matrix is screened according to the threshold interval, and the (X, Y) with a cell value meeting the threshold requirement is the target coordinate; the Boolean type matrix is screened according to the 0 / 1 matching, and the (X, Y) with a value of 1 is the target coordinate; the text encoding matrix first encodes the target semantic into a same-dimension vector, and then screens the coordinates according to the cosine similarity threshold;

[0029] When the unified determination standard is determined, the high similarity standards of the numerical type, Boolean type and text type matrix are classified into the same level, the standard difference is large, and the "strict and not loose" principle is adjusted, and the unique identification is associated with the target coordinates of each matrix, and the irrelevant coordinates are removed.

[0030] In a possible implementation, S4 specifically includes:

[0031] The weight is set based on the business demand or data importance, and the total weight is 1; in the e-commerce scenario, the click behavior matrix weight is 0.4, the product text description matrix weight is 0.3, the user collection matrix weight is 0.2, and the product numerical attribute matrix weight is 0.1.

[0032] First, normalize the original values of each matrix to the interval [0, 1], and then calculate the final fusion value according to "fusion value = Σ (matrix normalized value x corresponding weight)".

[0033] In a possible implementation, S4 specifically includes:

[0034] The similarity interval is divided into high similarity interval for fusion value ∈ [0.7, 1.0], medium similarity interval for ∈ [0.4, 0.7), and low similarity interval for ∈ [0, 0.4).

[0035] When locking the optimal selection, the high similarity interval result is preferentially extracted, the fusion value is sorted from high to low, the repeated results of the same unique identification are removed, the result with the highest fusion value is retained, and the S2 matrix core attribute is associated for auxiliary screening.

[0036] In a possible implementation, the S5 output result list is arranged in descending order of high->medium->low similarity level, each result item contains a subject identification, a similarity level, and a core matching basis, the subject identification includes a product ID, a name, and a core attribute; at the same time, the S2 matrix extension attribute and the user historical behavior are supplemented, the extension attribute includes logistics timeliness and praise rate, and the user historical behavior includes collection; a similarity level filtering and core dimension sorting tool is provided to help users quickly lock the optimal selection.

[0037] Compared with the prior art, the beneficial effects are:

[0038] 1. In the present scheme, the source data is split into independent data elements according to data attributes / business modules, and a two-dimensional matrix (numerical type / Boolean type / text encoding type) of the same type is constructed for each element, and the classification consistency check is performed to ensure that the matrix is buildable, non-redundant and covers all search instructions, realizes the conversion of data from "chaotic original state" to "structured and computable state", avoids the search deviation caused by mixed multi-type data, provides a standardized data basis for subsequent accurate similarity determination, and effectively improves the initial search accuracy.

[0039] 2. In this scheme, by mapping the user instruction to the corresponding matrix after parsing it into a standardized format, the similarity determination standards of different types of matrices are unified (such as numerical threshold, text similarity, and Boolean value level), and then the weight is allocated according to the business requirements to calculate the fusion value and divide the similar interval, realizing the collaborative retrieval of multi-dimensional data, avoiding the limitation of single dimension leading the result, and making the retrieval result reflect the user's demand for numerical, semantic, behavior and other aspects comprehensively;

[0040] 3. In this scheme, by ranking the high similarity results according to the fusion value and adding special visual identification, the matrix extension data (logistics time limit, good comment rate) and user historical behavior (already collected, viewed many times) are supplemented, and the core dimension comparison of multiple high similarity results is also provided, realizing the upgrade of the retrieval result from "simple list output" to "decision support", helping the user to quickly lock the optimal item. BRIEF DESCRIPTION OF DRAWINGS

[0041] The above description is only a summary of the technical scheme of the present application. In order to more clearly understand the technical means of the present application and can be implemented according to the content of the description, the following will be described in detail with reference to the preferred embodiments of the present application and with the help of the accompanying drawings.

[0042] Figure 1 The method steps flowchart of the present application. DETAILED DESCRIPTION

[0043] The preferred embodiments of the present application will be described in detail with reference to the accompanying drawings, but the present application can be realized in various different forms, and therefore the present application is not limited to the embodiments described below;

[0044] The technical scheme in the embodiments of the present application is to solve the problems in the above background art, and the general idea is as follows:

[0045] Embodiment: please refer to Figure 1 The present embodiment introduces a data retrieval method based on multi-element matrix fusion, including the following steps:

[0046] S1, multi-element classification, by defining classification dimensions, the whole set of source data is divided into multiple independent data elements according to the unified classification standard, and the classification granularity, i.e. the fineness of the classification, is controlled in the process of classification, and the consistency check is carried out to ensure that the granularity is reasonable, which establishes the basis for the construction of two-dimensional matrix for each element, specifically including:

[0047] S1.1, classification dimension definition, the core logic of defining classification dimension is that the dimension should adapt to the subsequent matrix construction (ensure that the data under the same dimension can form a structure unified matrix), and at the same time, it should match the user's possible retrieval instruction (ensure that the instruction can be quickly mapped to the matrix of the corresponding dimension);

[0048] According to the "attribute / type of data" (most commonly used, suitable for multi-type matrix): For example, the source data contains "user click times, product price, product description text, user collection status", and the classification standard is "data attribute". After classification, four "meta" are obtained, including: user click behavior data, product numerical attribute data, product text description data, and user collection behavior data.

[0049] According to the "business scenario / retrieval purpose" (fitting instruction requirements): For example, the retrieval scenario is "e-commerce product matching", and the source data is "user-related data, product-related data, and transaction-related data". The classification standard is "business module", and after classification, three "meta" are obtained, including: user behavior data, product attribute data, and transaction record data.

[0050] Each piece of data in the source data can find a corresponding "meta", such as "product praise rate" in e-commerce source data. If the "business module dimension" is selected, it should be classified into "evaluation data" meta and cannot miss any data. In addition, the same data cannot belong to two or more "meta" (without duplication). For example, "user clicks on product A record" can only be classified into "user behavior data" meta and cannot be classified into "product attribute data" meta.

[0051] S1.2, classification granularity control, granularity is the "coarseness" of classification, which should match the subsequent matrix purpose and retrieval accuracy. The classification granularity is determined by the priority of the retrieval target, and the specific control method is as follows:

[0052] First, according to the "core dimension rough division" (such as according to "data attribute" or "business module", divide into 2-3 coarse granularity meta), taking the e-commerce scenario as an example, it can be divided into three coarse granularity meta, including "user behavior data", "product attribute data", and "evaluation data";

[0053] Then, determine whether each coarse granularity meta is "data type uniform, can directly build matrix", if it can (such as "product attribute data" is all numerical / text, without mixing), keep the coarse granularity, and do not need to be further divided; if it cannot (such as "user behavior data" contains "click times", "whether to collect", and "stay time", which indicates that the data type is mixed and the matrix will be chaotic), therefore, the meta is "fine divided", and the data type / use of each meta is uniform.

[0054] Determine the medium granularity (final solution), make sure that "the data of each element can be built into a two-dimensional matrix with a unified structure" after disassembly, and the total number of matrices is ≤5 (not more than 8, otherwise the calculation amount will increase dramatically during fusion), the above product attribute data example is adjusted (medium granularity) and disassembled into 5 elements, including "click behavior data", "collection behavior data", "stay behavior data", "product attribute data" (numerical + text type, if the text proportion is high, it can be further disassembled into "product numerical attribute data" and "product text description data"), and "evaluation data" (text + numerical type).

[0055] S1.3, classification consistency check, after classification, use the following standards to check, if not meet, adjust to ensure reasonable granularity:

[0056] Each element's data can be "unified to build a two-dimensional matrix": the data type and purpose within the same element are consistent, such as "click behavior data" is all "user ID - product ID - click count", which can directly construct a numerical matrix of X=user ID, Y=product ID; If a element contains both numerical and text data (such as "product attribute data" has price and description), and both types of data have a high proportion, then it is disassembled into two elements (numerical attribute and text description), otherwise it cannot be built into a unified matrix;

[0057] No redundancy: the data use of the elements after disassembly does not overlap, such as "click behavior data" and "collection behavior data" do not overlap, and there is no need to disassemble "click + collection mixed data" elements;

[0058] Cover all search instructions: the search keywords that users may input can be matched to a certain element, such as searching for "click volume" corresponds to the "click behavior data" element, and searching for "collection status" corresponds to the "collection behavior data" element, and there is no situation where the instruction cannot find a corresponding element.

[0059] S2, matrix construction, based on the independent data elements obtained by S1 multi-element classification, build a two-dimensional matrix (X, Y) for each element, which has a unified structure and can be directly used for search matching;

[0060] S2.1, two-dimensional matrix (X, Y) axis definition, the core logic of axis definition is "use 'unique identifier' to define the subject, and use 'associated object / feature' to define the dimension", to ensure that each matrix cell can accurately correspond to the relationship of "who - associated with what - specific value", while adapting to the mapping needs of subsequent search instructions;

[0061] First, the X-axis or Y-axis must contain "unique identification", such as user ID, product ID, so as to ensure that each cell is positioned to the unique combination of "subject, associated object" (such as "user A, product B"); second, the dimension of the axis should be stable, and once determined, it should not be easily changed (such as product ID, attribute type), avoiding frequent reconstruction of the matrix; finally, the X-axis and Y-axis should cover all core subjects or features of the corresponding data elements, and there should be no omissions (such as the Y-axis of the product numerical attribute data, which should include all attributes defined by the element, such as price, sales, and inventory); taking the "user click behavior data" of S1 as an example, the X-axis is set as the user unique identification (user ID), covering all users with click behavior; the Y-axis is set as the product unique identification (product ID), covering all clicked products; each cell in the matrix corresponds to the "click times of a certain user on a certain product".

[0062] S2.2, the division standard of multi-type matrix, the core basis is "the type of matrix value + the logic of subsequent retrieval matching", to ensure that each matrix has a clear similarity determination method, and is completely adapted to the type of S1 data element, which is specifically divided into three categories:

[0063] Numerical matrix: the core distinguishing point is that the values in the matrix are all quantifiable numbers, which may be continuous (such as price 299.5 yuan) or discrete (such as click times 12 times). This kind of matrix is mainly used for "numerical interval retrieval", such as when the user searches for "price lower than 300 yuan" and "click volume ≥ 10 times", it is used for judgment; it is adapted to "user click behavior data" and "product numerical attribute data" in S1, and also to the numerical part of "evaluation data" (such as good comment rate and average score). The characteristics of this kind of matrix are that it supports threshold interval judgment and numerical similarity calculation (such as using Euclidean distance to calculate the difference between two numbers), and the range of values is very clear, such as click times ≥ 0, price ≥ 0, there will be no negative numbers;

[0064] Boolean matrix: the core distinguishing point is that the values in the matrix have only two options, 0 or 1, 0 represents "no corresponding relationship", and 1 represents "corresponding relationship"; this kind of matrix is mainly used for "whether class retrieval", such as when the user searches for "whether to collect a certain product" or "whether a certain product is free of charge"; it is adapted to "user collection behavior data" in S1, and also to "whether class" attributes in "product attribute data" (such as whether it is free of charge or whether it supports after-sales service); the logic of this kind of matrix is very simple, and when retrieving, only needs to determine "whether 0 or 1 matches", there is no intermediate value, so the retrieval efficiency is very high;

[0065] Text encoding matrix: the core distinguishing point is that the value in the matrix is not the original text, but the vector value of the text after encoding processing, such as TF-IDF value, Word2Vec word vector value, or BERT encoded floating point number; this kind of matrix is mainly used for "text semantic class retrieval", such as when the user searches for "find breathable T-shirt" "want to wear-resistant running shoes", the text part (such as user comments) in "product text description data" "evaluation data" in S1 is retrieved through semantic matching; this kind of matrix needs to do encoding processing on the original text first, and convert the text into a vector that can be calculated by the machine, and the value is usually a continuous floating point number (such as 0.82, -0.15), which supports semantic matching using cosine similarity and other methods;

[0066] It should be noted that the multi-type matrix division should comply with the "single type" principle, that is, each matrix can only be classified as one type, and cannot have both numerical and boolean values; at the same time, it needs to be "adapted to retrieval" to ensure that the matrix type can correspond to the user's possible retrieval instruction type, such as "semantic description type instruction" for text encoding matrix, to avoid the situation that "the instruction type and the matrix type do not match".

[0067] S2.3, matrix data filling rule, the filling rule mainly solves the problems of "data missing" and "data exception", ensures that the data in the matrix is complete and accurate, and can be directly used for subsequent retrieval calculation, and the filling logic of the same type of matrix should be unified to avoid contradiction;

[0068] Missing value filling, missing value is that a cell in the matrix is empty, such as "user A has not clicked on product B, and the click count is empty" "a product has no description text, and the text encoding value is empty"; the filling method of different types of matrices is different:

[0069] Numerical matrix: fill 0 first, such as user A has not clicked on product B, fill 0 for click count; if there is no data for the key attribute (such as product price), fill the median of the same type of product (such as the average price of a T-shirt is 250 yuan, and the price of a T-shirt is missing, fill 250 yuan), to avoid affecting the overall numerical distribution;

[0070] Boolean matrix: fill 0 uniformly, such as user A has not collected product B, fill 0 to represent "not collected"; if a product has no "whether to ship" information, fill 0 to represent "default not shipped";

[0071] Text encoding matrix: if there is no text data (such as no description for a product), fill a full 0 vector, which means "no text features", and subsequent retrieval will not be misjudged as having semantic correlation;

[0072] Abnormal value processing, abnormal value is data that is obviously not logical, and its processing is divided into two steps:

[0073] Outlier determination and processing: In numerical matrices, data exceeding 3 times the standard deviation (such as the standard deviation of the price of similar goods being 50 yuan, 3 times being 150 yuan, and exceeding 300+150=450 yuan being an outlier) or logically inconsistent data (such as negative click counts or 0 yuan prices but not gifts) are considered outliers. For numerical matrices, outliers are replaced with 3 times the standard deviation of the column (i.e., the attribute corresponding to the Y-axis), such as the 450 yuan boundary value mentioned above, and an abnormal price of 10,000 yuan is replaced with 450 yuan. Boolean and text encoding matrices are less likely to have outliers, and if non-0 / 1 Boolean values or invalid vector values appear, they are filled with default values (0 or all-0 vectors).

[0074] S3, Similar interval determination, based on the multi-type matrix constructed in S2, through the logic of "instruction to condition, matching corresponding matrix, positioning target coordinates, and unified determination standard", the similar data meeting the user's search requirements are filtered out from the matrix, providing accurate input for subsequent multi-matrix data fusion;

[0075] S3.1, Instruction analysis and standardization, first split the core keywords in the instruction (including search subject, demand type, and limiting conditions), then match the corresponding determination logic according to the S2 matrix type (numerical, Boolean, and text encoding), and finally convert it to a unified format;

[0076] Keyword splitting: Extract the "search subject" (such as goods, users), "core demand" (such as price below 300, breathable, whether collected), and "limiting conditions" (such as the last 7 days, similar goods) from the instruction. For example, the user instruction "find T-shirts with click count ≥10 and breathable in the last 7 days" is split into: search subject = T-shirt, core demand = click count ≥10 (numerical), breathable (text semantic), and limiting condition = the last 7 days.

[0077] Demand type matching: According to the keywords corresponding to the S2 matrix type, determine the determination method: numerical type demand (such as price, click count) corresponds to "threshold interval determination", whether type demand (such as whether collected, whether free shipping) corresponds to "0 / 1 matching determination", and semantic type demand (such as breathable, wear-resistant) corresponds to "similarity threshold determination".

[0078] Standardized format conversion: Convert the analysis result to a unified format of "{search subject}+{dimension}+{operator}+{target value / threshold value}"; for example, the above instruction is standardized as: {search subject = T-shirt}, {dimension = click count, operator = ≥, target value = 10, limiting condition = time = the last 7 days}, {dimension = product description semantics, operator = ≥, threshold value = 0.8 (cosine similarity), target semantics = breathable}.

[0079] S3.2, Instruction-matrix mapping, according to the standardized search conditions, match to one or more matrices corresponding to S2, ensure that the search is only in the relevant matrix, the core is to correspond the "dimension type" of the standardized condition to the matrix type of S2, correspond the "search subject / dimension" to the axis definition (X / Y axis) of S2 matrix, realize the accurate mapping of "one condition corresponds to one matrix"; The specific mapping rules are as follows:

[0080] Numerical conditions (such as price ≤ 300, click volume ≥ 10): mapped to the "numerical matrix" of S2, and the axis dimension of the matrix needs to contain the search subject in the condition (such as the price condition mapped to the "commodity numerical attribute matrix", because the matrix X axis = commodity ID, Y axis = price);

[0081] Boolean conditions (such as whether to collect, whether to ship): mapped to the "Boolean matrix" of S2, for example, "whether to collect" is mapped to the "user collection behavior matrix" (X axis = user ID, Y axis = commodity ID);

[0082] Text semantic conditions (such as breathable, wear-resistant): mapped to the "text encoding matrix" of S2, for example, "breathable" is mapped to the "commodity text description matrix" (X axis = commodity ID, Y axis = text feature dimension);

[0083] Multi-condition combination (such as "click volume ≥ 10 and breathable"): map multiple corresponding matrices respectively to form a "matrix combination list" (such as the numerical click behavior matrix + the text encoding commodity description matrix);

[0084] Note that a standardized condition only maps one matrix to avoid repeated searches; If there is no corresponding matrix for the condition (such as searching for "logistics time limit" but no related matrix is constructed), prompt "no matching data source" and feedback to the S1 classification link for optimization; The matrix mapped by multiple conditions needs to contain the same "unique identifier" (such as commodity ID) to ensure that the coordinates can be aligned later (such as the click behavior matrix and the commodity description matrix both take commodity ID as the core identifier).

[0085] S3.3, Matrix coordinate positioning, according to the matrix type, use the corresponding similarity judgment method to select the coordinates whose cell values meet the conditions, ensure that each positioned coordinate meets the "similarity" requirement, and the type positioning operation is as follows:

[0086] Numerical matrix (such as click behavior matrix, commodity numerical attribute matrix): positioned according to "threshold interval", for example, the condition "click volume ≥ 10" selects all "(user ID, commodity ID) coordinates" whose cell values are ≥ 10 in the click behavior matrix; The condition "price ≤ 300" selects the "(commodity ID, price) coordinates" whose cell values are ≤ 300 in the commodity numerical attribute matrix, and records the corresponding numerical value;

[0087] Boolean matrix (such as user favorite behavior matrix): locate by "0 / 1 matching", for example, the condition "whether favorite = yes" filters out the (user ID, product ID) coordinates with cell value = 1; the condition "whether free shipping = yes" filters out the (product ID, whether free shipping) coordinates with cell value = 1;

[0088] Text encoding matrix (such as product text description matrix): first encode the target semantics (such as "breathable") in the standardized condition into a vector of the same dimension as the matrix, and then locate by "semantic similarity threshold" - calculate the cosine similarity between the vector of each (product ID, text feature dimension) coordinate in the matrix and the target vector, filter out the coordinates with similarity ≥ the set threshold (such as 0.8), and record the corresponding similarity value;

[0089] Output of positioning results: each matrix outputs "list of coordinates meeting the conditions + corresponding cell values", in the format "matrix name: [(X1, Y1): value1, (X2, Y2): value2...]". For example, "click behavior matrix: [(user A, product B): 12, (user C, product D): 15]".

[0090] S3.4, interval consistency alignment, unify the "similar interval" judgment standards of multiple matrices, and ensure that the coordinates located are related, so that the subsequent fusion is the similar data of "the same subject / object"; the specific alignment operation is as follows:

[0091] Standard consistency alignment: unify the "similar strength standards" of different matrices, for example, "high similarity" (click volume ≥ 10) of numerical type, "high similarity" (similarity ≥ 0.8) of text type, and "high similarity" (value = 1) of Boolean type are classified as the same level of similarity standard; if the threshold values of multiple conditions differ greatly (such as numerical threshold is loose and text threshold is strict), adjust according to the "take strict and not take loose" principle (such as reduce the text similarity threshold to 0.7 or increase the numerical threshold to 15) to ensure the matching of similarity degree;

[0092] Coordinate correlation alignment: take "common unique identifier" (such as product ID, user ID) as the link to correlate the positioning results of multiple matrices, for example, (user A, product B) located by the click behavior matrix and (product B, text feature) located by the product description matrix are aligned through the common identifier "product B", and the intersection coordinates (i.e. only the coordinates corresponding to the product ID / user ID contained in multiple matrices are retained) are retained;

[0093] Redundant coordinate elimination: eliminate irrelevant coordinates (such as product E in the click behavior matrix does not meet the semantic similarity threshold in the product description matrix), and only retain the coordinates that meet the conditions of "all mapping matrices", to ensure that the data fused subsequently are all similar data of full dimension.

[0094] S4, fusion computing layer, based on S3 similarity interval determination of aligned multi-matrix data, through weighted fusion, similarity interval division logic, multi-dimensional similar data is integrated into a unified search result, the core is to realize "multi-dimensional data collaborative support search conclusion", both to ensure the accuracy of the results, and let the user clearly know the result source and basis;

[0095] S4.1, weighted fusion, each aligned matrix data is assigned a weight, the final fusion value is calculated by "weighted sum", the weight proportion is set according to business needs or data importance;

[0096] Weight assignment: set the weight of each matrix according to the priority of the demand (the sum of the weights is 1), for example, in the e-commerce scenario, the weight of "click behavior matrix (numeric type)" is 0.4, the weight of "product text description matrix (text coded type)" is 0.3, the weight of "user collection matrix (Boolean type)" is 0.2, and the weight of "product numerical attribute matrix (numeric type)" is 0.1;

[0097] Data normalization: map the original values of different matrices to the [0, 1] interval (avoid the influence of numerical range difference on the result), for example, the numerical value "clicks 15 times" (maximum value 100) is normalized to 0.15, the text value "similarity 0.8" is directly used, the Boolean value "value 1" is normalized to 1, and the Boolean value "value 0" is normalized to 0;

[0098] Weighted calculation: fusion value = Σ (certain matrix normalized value x corresponding weight), for example, the fusion value of a certain product = (0.15 x 0.4) + (0.8 x 0.3) + (1 x 0.2) + (0.7 x 0.1) = 0.06 + 0.24 + 0.2 + 0.07 = 0.57.

[0099] S4.2, similarity interval division, combined with the weight setting of the e-commerce matrix in S4.1 (click 0.4, text 0.3, collection 0.2, price 0.1), according to "fusion value high and low", corresponding "multi-dimensional matching degree" is divided, as follows:

[0100] High similarity interval: fusion value ∈ [0.7, 1.0] business meaning: at least 3 or more matrix data highly meet the demand, the core dimension (click, text) is outstanding; for example, the text semantic similarity is ≥0.8 (normalized 0.8), the click value is normalized ≥0.6, and the collection state is 1 (normalized 1), the fusion value can reach (0.6 x 0.4) + (0.8 x 0.3) + (1 x 0.2) + (0.9 x 0.1) = 0.24 + 0.24 + 0.2 + 0.09 = 0.77, which belongs to high similarity, representing "high matching between search results and demand";

[0101] Middle similarity interval: fusion value ∈ [0.4, 0.7) Business meaning: 2-3 matrix data meet the demand, core dimension meets the standard but auxiliary dimension has short board; For example, text semantic similarity 0.7 (0.7 after normalization), click volume normalized value 0.5, collection state 0 (0 after normalization), fusion value is (0.5x0.4) + (0.7x0.3) + (0x0.2) + (0.8x0.1) = 0.2 + 0.21 + 0 + 0.08 = 0.49, belongs to middle similarity, represents "search results basically meet the demand, some details need to be considered";

[0102] Low similarity interval: fusion value ∈ [0, 0.4) Business meaning: only 1 or 0 matrix data meets the demand, core dimension does not meet the standard; For example, text semantic similarity 0.5 (0.5 after normalization), click volume normalized value 0.3, collection state 0, fusion value is (0.3x0.4) + (0.5x0.3) + (0x0.2) + (0.6x0.1) = 0.12 + 0.15 + 0 + 0.06 = 0.33, belongs to low similarity, represents "search results have low matching degree with demand, not recommended to be selected preferentially".

[0103] S4.3, high similarity result presentation, for search results falling into S4.2 high similarity interval, quickly lock the optimal choice, high similarity result is the "optimal matching item" of user demand, the presentation needs to be carried out around "quick identification, clear basis, convenient selection", the core is to convert S2 matrix data and S4.1 fusion value into information that can be directly understood by users, avoiding technical data stacking, which is as follows:

[0104] Sorting and visual focusing:

[0105] Priority sorting: in the search result list, all high similarity results are placed at the top, and are sorted again according to the "fusion value from high to low" of S4.1 (such as fusion value 0.92 result > 0.85 result > 0.72 result), the sorting basis directly follows the calculation result of weighted fusion, ensuring that "the higher the fusion value, the closer to the front", which meets the user's expectation of "more matching" result;

[0106] Exclusive visual identification: add differentiated label (such as "[high matching recommendation]") to high similarity result, the label color is different from that of middle and low similarity results (such as red / orange), and the "core matching dimension" is highlighted in the result card, for example, "semantic similarity 0.88" of S2 "commodity text description matrix" (expressed as "description and demand fit degree 88%"), "click volume 92 times" of S2 "click behavior matrix" (expressed as "click volume exceeds 90% of similar commodities"), so that users can instantly perceive the core basis of "high similarity";

[0107] Simplify redundant information: The result card only retains attributes strongly related to core needs, all from S2 matrix - such as product name (S2 "product numerical attribute matrix" x-axis product ID associated name), price (S2 "product numerical attribute matrix" y-axis price dimension), inventory (S2 "product numerical attribute matrix" y-axis inventory dimension), avoid adding historical browsing records, irrelevant labels and other interference information, help users quickly confirm "whether it meets the basic preferences".

[0108] Core details supplement:

[0109] Highly similar results may meet core needs, but users may be concerned about "implicit needs" (such as logistics, after-sales, real evaluation), which require the extension of S2 matrix data, and the data must be consistent with the previous classification logic (from S1 classification metadata):

[0110] Supplement attributes related to S2 matrix: Extract "logistics timeliness (such as '24 hours after ordering') "after-sales policy (such as '7-day no reason for return and exchange') from S2 "product numerical attribute matrix" and "good rate (such as 98%) "core evaluation keywords (such as 'breathable and not sweaty''soft fabric') from S2 "evaluation data matrix" (S1 classification "evaluation data" meta corresponding matrix), these data come from independent metadata after S1 classification, ensuring consistency with the search dimension;

[0111] Label "user historical association behavior": If the highly similar result is "collected" (value = 1) in S2 "user collection matrix", or has "multiple click records" (such as click times ≥ 3 times) in S2 "click behavior matrix", then prompt "You have collected the product" "You have viewed the product 4 times" in the result card, associate user past behavior to enhance trust and reduce decision-making hesitation cost.

[0112] Multiple result comparison:

[0113] If there are 2 or more highly similar results, provide "core dimension comparison tool", comparison dimensions come from S2 matrix and S4.1 fusion dimension, avoid manual switching card comparison:

[0114] Comparison dimension setting: Default display "fusion value, semantic similarity, click volume, price, inventory" 5 core dimensions, respectively corresponding to S4.1 fusion value, S2 "product text description matrix", S2 "click behavior matrix", S2 "product numerical attribute matrix";

[0115] Comparison form: Use "word points + key data highlight" to present, for example:

[0116] Result A (fusion value 0.88): semantic similarity 0.9 (high), click volume 85 times (high), price 269 yuan, inventory 60 pieces;

[0117] Result B (fusion value 0.82): semantic similarity 0.85 (high), clicks 72 (medium), price 249 yuan (low), inventory 120 items;

[0118] Preference guidance: Based on the user's search instructions (such as "find a breathable T-shirt with a good price-performance ratio"), highlight the "price" dimension in the comparison results (such as "249 yuan" in result B highlighted in blue), guide users to focus on differences that are more relevant to their own needs, and improve comparison efficiency.

[0119] S5. Retrieval Presentation: Based on the entire process of S1-S4.3 (multi-class classification → matrix construction → similarity interval determination → fusion calculation → high similarity presentation and optimization), the final output is a list of results that can be used for direct decision-making. The list is the core of the final output. Each result item is sorted by similarity level from "high → medium → low" (high similarity is placed at the top) and includes "subject identifier + similarity level + core matching criteria". All data is traced back to the S2 matrix and S4 calculation results, with no redundant information. Taking the e-commerce "breathable T-shirt retrieval" scenario as an example:

[0120] High similarity result 1 (fusion value 0.88)

[0121] Main identifiers: Product ID (P6789), Product Name (Men's Ice Silk Breathable T-shirt), Core Attributes (White, Size L, ¥269, 60 pieces in stock)

[0122] Similarity rating: [High match recommendation] (fusion value 0.88, falls into the S4.2 high similarity interval [0.7, 1.0])

[0123] Core matching criteria:

[0124] Text description: Semantic similarity with "breathable T-shirt" is 0.9 (from S2 product text description matrix, normalized to 0.9, weight 0.3, contribution value 0.27).

[0125] Click count: 92 times (S2 click behavior matrix, maximum value 100 times, normalization 0.92, weight 0.4, contribution value 0.368).

[0126] Collection status: Collectiond (S2 user collection matrix, normalized to 1, weight 0.2, contribution value 0.2);

[0127] Price: 269 yuan (meets "≤300 yuan", S2 product numerical attribute matrix, normalization 0.8, weight 0.1, contribution value 0.08).

[0128] Additional note: You have viewed this product 4 times. It is currently in stock and will be shipped within 24 hours after you place your order (from S2 matrix supplementary attributes and user history behavior).

[0129] High similarity result 2 (fusion value 0.82)

[0130] Body identification: commodity ID (P8921), commodity name (men's pure cotton breathable T-shirt), core attribute (black, M size, 249 yuan, 120 pieces in stock)

[0131] Similarity level: [high matching recommendation] (fusion value 0.82, falling into S4.2 high similarity interval)

[0132] Core matching basis:

[0133] Text description similarity 0.85 (S2 text matrix, contribution value 0.255);

[0134] Clicks: 72 times (S2 click matrix, normalized 0.72, contribution value 0.288);

[0135] Collection status: not collected (contribution value 0);

[0136] Price: 249 yuan (normalized 0.92, contribution value 0.092);

[0137] Additional prompt: good comment rate 98%, core evaluation "breathable and not stuffy" (from S2 evaluation matrix).

[0138] Medium similarity result (fusion value 0.57), low similarity result (fusion value 0.33)

[0139] The presentation logic is consistent with the high similarity, but only "body identification + similarity level + key matching basis" are retained (such as medium similarity result only shows "text similarity 0.8 + price 299 yuan, low click volume 15 times"), and the vision is weakened (such as gray label "basically consistent" "low matching"), to avoid interfering with the selection of high similarity results.

[0140] Finally, it should be noted that: obviously, the above embodiments are only examples for clearly illustrating the present application, and are not limited to the implementation. For ordinary skilled persons in the art, on the basis of the above description, other different forms of changes or variations can also be made. Here, it is not necessary and impossible to enumerate all the implementation modes. The obvious changes or variations derived therefrom are still within the protection scope of the present application.

Claims

1. A data retrieval method based on multi-element matrix fusion, characterized in that, Comprise the following steps: S1, multi-element classification, the whole set of source data is split into multiple independent data elements according to unified classification standard by defining classification dimension, the classification process should control the classification granularity, that is, the degree of classification, and ensure that the granularity is reasonable through consistency check, the consistency check includes: the data type and purpose in each data element are consistent, can construct two-dimensional matrix uniformly; there is no redundancy between data elements; all user input search keywords can be corresponded to at least one data element; S2, matrix construction, based on the independent data elements obtained by S1 multi-element classification, a two-dimensional matrix (X, Y) with unified structure and directly used for search matching is built for each element, wherein the X axis or Y axis contains a unique identifier, the matrix is divided into numerical type, Boolean type or text encoding type according to the type of value, and each matrix belongs to only one type; S3, similar interval determination, based on the multi-type matrix constructed in S2, the user search instruction is parsed into core keywords and converted into standardized search conditions, the corresponding matrix is matched according to the condition type, the coordinates satisfying the conditions in the matrix are located, the similar determination standards of different matrices are unified, and the target coordinates of each matrix are associated based on the unique identifier; The user search instruction is parsed into a standard format "{search subject}+{dimension}+{operator}+{target value / threshold}", and the corresponding matrix is matched according to the dimension type: numerical type condition maps numerical type matrix, Boolean type condition maps Boolean type matrix, and text semantic type condition maps text encoding matrix; S4, fusion calculation layer, based on the aligned multi-matrix data of S3 similar interval determination, the final fusion value is calculated by assigning weight to each aligned matrix data, the matching degree of multiple dimensions is divided according to the fusion value, the similarity interval is divided, and the optimal choice is quickly locked for the search results of the similarity interval division; The weight is set based on business requirements and the sum is 1, the original value of each matrix is normalized to the interval [0, 1], and the fusion value = Σ (normalized value x corresponding weight); The similarity interval division is: high similarity [0.7, 1.0], medium similarity [0.4, 0.7], and low similarity [0, 0.4]; S5, search presentation, based on the whole process of S1 to S4, the final output is a result list that can be used for direct decision.

2. The data retrieval method based on multi-element matrix fusion according to claim 1, characterized in that, The specific process of controlling classification granularity in S1 is: first, split the source data into 2-3 coarse granularity data elements according to the core dimension, then judge whether the data type of each coarse granularity element is uniform and can be directly used to construct a two-dimensional matrix, if not, perform fine splitting until the data type and purpose of each data element are uniform, and the total number of final data elements does not exceed 8.

3. The data retrieval method based on multi-element matrix fusion according to claim 1, characterized in that, In S2, if the two problems of "data missing" and "data abnormality" occur, the matrix data filling step is performed to ensure that the data in the matrix is complete and accurate and can be directly used for subsequent search calculation, as follows: In the missing value processing, the numerical type matrix is filled with 0 first, the missing key attribute is filled with the median of the same type data; the Boolean type matrix is uniformly filled with 0; the text encoding matrix is filled with a full 0 vector; In the process of abnormal value, numerical matrix beyond 3 times the standard deviation or logical contradiction data, replace the corresponding column 3 times the standard deviation boundary value; Boolean, text encoding matrix appears non-standard value, fill 0 or all 0 vector.

4. The data retrieval method based on multi-element matrix fusion according to claim 1, characterized in that, S5 output result list is arranged in descending order of high→medium→low similarity level, each result item contains subject identification, similarity level, core matching basis, and the subject identification includes commodity ID, name and core attribute; meanwhile, S2 matrix extension attribute and user historical behavior are supplemented, the extension attribute includes logistics time limit, praise rate, and the user historical behavior includes collection; similarity level filtering and core dimension sorting tools are provided to help users quickly lock the optimal choice.

Images