A private domain e-commerce data search method and system based on big data

Abstract

The present invention relates to the field of private domain e-commerce technology, and specifically discloses a private domain e-commerce data search method and system based on big data. The present invention first collects user behavior characteristics (search, browsing time, product click volume) and purchase characteristics (repurchase cycle, time preference), constructs style preference models and purchase preference models, quantifies users' attention to product styles based on browsing data, and performs semantic expansion based on the explicit semantics of search keywords to capture users' potential needs. At the same time, segmented product quantization technology is used to compress the extended semantics and purchase preferences. A double-layer compact index is used to achieve parallel retrieval of explicit and associated needs. Finally, the structured search results are visualized, integrating basic product information and style extension features to meet users' potential extension intentions and avoid the need for users to conduct multiple repeated searches.

Description

Technical Field

[0001] The present invention relates to the field of private domain e-commerce technology, and in particular to a private domain e-commerce data search method and system based on big data. Background Art

[0002] Private domain e-commerce refers to a business model that relies on an individual or company's own traffic pool (such as WeChat official accounts, mini-programs, social groups, and apps) to sell goods and provide services. Unlike traditional e-commerce platforms (such as Taobao and JD.com), private domain e-commerce emphasizes establishing and maintaining direct connections with consumers, forming a stable customer base, and thus maximizing long-term value.

[0003] Existing private domain e-commerce platforms usually rely on basic keywords for search, and this search method only returns results based on literal matching, which ignores the user's potential intentions (for example, when a user searches for "summer dress", the traditional method only returns results based on literal matching, ignoring the user's potential intentions, such as "breathable fabric" and "vacation style"). It cannot provide accurate related information (breathable fabric, etc.), which in turn causes users to need to repeat searches multiple times. A private domain e-commerce data search method based on big data is needed to solve the above problems. Summary of the Invention

[0004] The purpose of the present invention is to provide a private domain e-commerce data search method and system based on big data to solve the technical problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A private domain e-commerce data search method based on big data, comprising:

[0007] Obtain user search feature information on the private domain e-commerce platform, where the user search feature information includes behavioral feature information and purchase feature information;

[0008] Obtaining browsing characteristic information and search keywords based on the behavioral characteristic information, obtaining multiple browsed product click counts and multiple browse click dwell times based on the browsing characteristic information, and obtaining style preference characteristics based on the multiple browsed product click counts and multiple browse click dwell times;

[0009] Performing semantic extraction on the search keyword to obtain explicit semantic features, and obtaining extended semantic features based on the explicit semantic features and style preference features;

[0010] Obtaining repeat purchase data information and timeliness information based on the purchased product characteristic information, and obtaining purchase preference characteristics based on the repeat purchase data information and timeliness information;

[0011] Performing segmented product quantization compression on the extended semantic features and the purchase preference features to obtain compressed semantic association features, and simultaneously performing double-layer compact indexing on the explicit semantic features and the compressed semantic association features to obtain explicit results and association results;

[0012] The private domain e-commerce search results are visualized based on the explicit results and the associated results.

[0013] Preferably, the step of obtaining style preference features based on the multiple click times of the browsed products and the multiple browse click dwell times includes:

[0014] sorting the plurality of browsed products by the number of clicks in descending order to obtain a browsed product click number sorting table;

[0015] Sort the plurality of browsing click and stay times in descending order to obtain a browsing click and stay time sorting table;

[0016] Obtaining the first product corresponding to the first product click count according to the browsed product click count ranking table;

[0017] Obtaining the second product corresponding to the first product in the browsing click and dwell time ranking table according to the browsing click and dwell time ranking table, and subjecting the second product and the first product to dynamic load coefficient constraints to obtain a balanced cluster set, and performing centroid screening on the balanced cluster set based on the K-means clustering algorithm to obtain the product centroid;

[0018] Performing segmented semantic feature screening on the product centroid based on an attention mechanism to obtain multiple segmented semantic features;

[0019] Perform keyword extraction on the plurality of segmented semantic features based on the BERT language model to obtain a first keyword, and capture contextual associations of the first keyword based on a multi-head self-attention mechanism to obtain a keyword semantic text;

[0020] The style preference features of the keyword semantic text are extracted based on a preset text library.

[0021] Preferably, the step of performing semantic extraction on the search keyword to obtain explicit semantic features, and obtaining extended semantic features based on the explicit semantic features and style preference features includes:

[0022] Performing word segmentation processing on the search keyword to obtain multiple independent words of the search keyword, converting the multiple independent words into word vectors based on the Word2Vec model to obtain independent word set vectors, and defining the semantic features corresponding to the independent word set vectors as explicit semantic features;

[0023] Mapping the style preference features into style semantic vectors aligned with the independent word set vector dimensions through a fully connected layer;

[0024] Comparing and screening the independent word set vector with the style semantic vector based on a cosine similarity model to obtain an independent word-style semantic subset vector;

[0025] Performing vector splicing on the independent word-style semantic subset vector and the style semantic vector based on a preset time sequence to obtain a spliced ​​vector;

[0026] The splicing vector is input into a preset bidirectional LSTM network for contextual semantic enhancement to obtain an enhanced splicing vector, and the semantic features of the enhanced splicing vector are used as extended semantic features.

[0027] Preferably, the step of obtaining repeat purchase data information and timeliness information based on the purchased commodity characteristic information, and obtaining purchase preference characteristics based on the repeat purchase data information and timeliness information includes:

[0028] Obtaining historical purchase timestamps and multiple product categories based on the purchased product feature information, and obtaining the time interval between two adjacent purchases under each of the same product categories;

[0029] Marking multiple behaviors with time intervals less than a preset threshold as repeat purchase events, and defining data information corresponding to the repeat purchase events as repeat purchase data information;

[0030] Obtaining a time sliding window based on the historical purchase timestamps, and obtaining a seasonal distribution interval based on the time sliding window, dividing the plurality of product categories into the seasonal distribution intervals to obtain seasonal interval naturalized product categories, and using time information of the seasonal interval naturalized product categories in the seasonal distribution intervals as timeliness information;

[0031] Performing time-efficiency screening on the repurchase data information according to the time-efficiency information to obtain screened repurchase data information, and encoding the screened repurchase data information to obtain a screened repurchase data code;

[0032] Based on the Word2Vec model, the filtered repurchase data encoding is converted into a word vector to obtain the filtered repurchase data encoding vector, and the filtered repurchase data encoding vector is used as the purchase preference feature.

[0033] Preferably, the step of performing piecewise product quantization compression on the extended semantic features and the purchase preference features to obtain compressed semantic association features comprises:

[0034] Mapping the extended semantic features into an extended semantic feature vector;

[0035] Mapping the purchase preference feature into a purchase preference feature vector, and concatenating the extended semantic feature vector and the purchase preference feature vector to obtain a joint vector;

[0036] Dividing the joint vector into a plurality of sub-vector segments according to a preset length, and clustering each sub-vector segment based on K-means clustering to obtain a plurality of cluster centroids, wherein each cluster centroid represents an index;

[0037] The plurality of cluster centroids are mapped into the plurality of cluster centroid vectors, and the plurality of cluster centroid vectors are multiplied and quantized to obtain particle vector values, and the semantics corresponding to the particle vector values ​​are used as compressed semantic association features.

[0038] Preferably, the step of simultaneously performing double-layer compact indexing on the explicit semantic features and the compressed semantic association features to obtain explicit results and association results includes:

[0039] The explicit semantic feature is constructed based on the explicit feature inverted index to obtain a first-level index, and the first-level index is mapped to a hash bucket based on locality sensitive hashing (LSH) to obtain a hash first-level index, and a result corresponding to the hash first-level index is used as the explicit result;

[0040] constructing a second-layer index from the compressed semantic association features based on a compressed feature graph index, and constructing the second-layer index into a multi-layer graph index based on a hierarchical navigable small-world (HNSW) graph algorithm, wherein each layer of the graph index corresponds to a product identifier;

[0041] The result corresponding to the multi-layer graph index is used as the association result.

[0042] This application also provides a private domain e-commerce data search system based on big data, including:

[0043] The first acquisition module is used to obtain user search feature information of the private domain e-commerce platform, wherein the user search feature information includes behavioral feature information and purchase feature information;

[0044] A second acquisition module is configured to acquire browsing characteristic information and search keywords based on the behavioral characteristic information, and acquire multiple browsed product click counts and multiple browsed click dwell times based on the browsing characteristic information, and acquire style preference characteristics based on the multiple browsed product click counts and multiple browsed click dwell times;

[0045] An extraction module, configured to perform semantic extraction on the search keyword to obtain explicit semantic features, and to obtain extended semantic features based on the explicit semantic features and style preference features;

[0046] A third acquisition module is configured to acquire repeat purchase data information and timeliness information based on the purchased commodity characteristic information, and acquire purchase preference characteristics based on the repeat purchase data information and timeliness information;

[0047] a fourth acquisition module, configured to perform piecewise product quantization compression on the extended semantic features and the purchase preference features to obtain compressed semantic association features, and to simultaneously perform double-layer compact indexing on the explicit semantic features and the compressed semantic association features to obtain explicit results and association results;

[0048] A presentation module is used to visually present the private domain e-commerce search results based on the explicit results and the associated results.

[0049] Preferably, the second acquisition module includes:

[0050] A first sorting unit is used to sort the click counts of the browsed products in descending order to obtain a sorting table of click counts of browsed products;

[0051] A second sorting unit is used to sort the plurality of browsing click and stay times in descending order to obtain a browsing click and stay time sorting table;

[0052] A first acquiring unit is configured to acquire, according to the browsed product click count ranking table, a first product corresponding to a first browsed product with a highest number of clicks;

[0053] A second acquisition unit is configured to acquire a second product corresponding to a first product in the browsing click and dwell time ranking table according to the browsing click and dwell time ranking table, and subject the second product and the first product to dynamic load coefficient constraints to obtain a balanced cluster set, and perform centroid screening on the balanced cluster set based on a K-means clustering algorithm to obtain a product centroid;

[0054] A screening unit, configured to screen the product centroid for segmented semantic features based on an attention mechanism to obtain multiple segmented semantic features;

[0055] A first extraction unit is configured to perform keyword extraction on the plurality of segmented semantic features based on a BERT language model to obtain a first keyword, and to capture contextual associations of the first keyword based on a multi-head self-attention mechanism to obtain a keyword semantic text;

[0056] The second extraction unit is configured to extract the style preference features of the keyword semantic text based on a preset text library.

[0057] The present application also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0058] The present application also provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when executed by a processor.

[0059] The beneficial effects of the present application are as follows: the present invention first collects user behavior characteristics (search, browsing time, product click volume) and purchase characteristics (repurchase cycle, time preference), constructs a style preference model and a purchase preference model, quantifies the user's attention to the product style based on the browsing data, and performs semantic expansion in combination with the explicit semantics of the search keywords to capture potential demands such as "breathable fabrics" and "vacation style". At the same time, the segmented product quantization technology is used to compress the extended semantics and purchase preferences, and the parallel retrieval of explicit demands and associated demands is realized through a double-layer compact index. Finally, the structured search results are visualized, and the basic information of the product and the style extension features are integrated to meet the user's potential intention needs, avoiding the need for users to repeat searches multiple times. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a schematic diagram of a method flow chart according to an embodiment of the present application.

[0061] Figure 2 This is a schematic diagram of the system structure of an embodiment of the present application.

[0062] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of the present application.

[0063] The realization of the objectives, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0064] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0065] like Figure 1-3 As shown, this application provides a private domain e-commerce data search method based on big data, including:

[0066] S1. Obtain user search feature information on the private domain e-commerce platform, where the user search feature information includes behavioral feature information and purchase feature information;

[0067] S2. Obtaining browsing characteristic information and search keywords based on the behavioral characteristic information, obtaining multiple browsed product click counts and multiple browse click dwell times based on the browsing characteristic information, and obtaining style preference characteristics based on the multiple browsed product click counts and multiple browse click dwell times;

[0068] S3. Performing semantic extraction on the search keyword to obtain explicit semantic features, and obtaining extended semantic features based on the explicit semantic features and style preference features;

[0069] S4. Obtaining repeat purchase data information and timeliness information based on the purchased product characteristic information, and obtaining purchase preference characteristics based on the repeat purchase data information and timeliness information;

[0070] S5. Performing piecewise product quantization compression on the extended semantic features and the purchase preference features to obtain compressed semantic association features, and simultaneously performing double-layer compact indexing on the explicit semantic features and the compressed semantic association features to obtain explicit results and association results;

[0071] S6. Visually present the private domain e-commerce search results based on the explicit results and the associated results.

[0072] As described in the above steps S1-S6, since existing private domain e-commerce platforms usually rely on basic keywords for searching, and this search method only returns results based on literal matching, it will ignore the user's potential intentions (for example, when a user searches for "summer dress", the traditional method only returns results based on literal matching, ignoring the user's potential intentions, such as "breathable fabrics", "vacation style", etc.), and cannot provide accurate related information (breathable fabrics, etc.), which in turn causes the user to need to repeat the search multiple times. The present invention first obtains the user search feature information of the private domain e-commerce platform, wherein the user search feature information includes behavioral feature information and purchase product feature information. In this way, by collecting the user's behavior and purchase-related information on the private domain e-commerce platform, the foundation is laid for the subsequent in-depth analysis of user needs and preferences. By acquiring this feature information, we can comprehensively understand user activities on the platform, including search, browsing, and purchasing behaviors, thereby mining users' potential needs. Browsing feature information and search keywords are then obtained based on the behavioral feature information. Multiple click counts and dwell times of browsed items are then obtained based on the browsing feature information. Style preference features are then obtained based on the multiple click counts and dwell times of browsed items. Browsing and search-related data is extracted from the behavioral feature information, and further analysis is performed to obtain users' style preference features. This helps more accurately grasp users' preferences and provides guidance for expanding and optimizing search results. By analyzing the click counts and dwell times of browsed items, we can understand the user's level of interest in different product styles and thus infer their style preferences. Furthermore, users' browsing and search behaviors can intuitively reflect their interests. By mining this behavioral data, we can discover users' potential style preferences and make search results more consistent with their expectations. Semantic extraction is then performed on the search keywords to obtain explicit semantic features. Extended semantic features are then obtained based on the explicit semantic features and style preference features. By mining the explicit semantic features of the search keywords and combining them with the extended semantic features of the style preference features, search results are made more comprehensive and more aligned with users' potential intentions. It can not only match the literal meaning, but also take into account the related semantics that the user may be interested in, thereby improving the accuracy and relevance of the search. Secondly, it obtains the repeat purchase data information and timeliness information based on the purchased product feature information, and obtains the purchase preference characteristics based on the repeat purchase data information and timeliness information. By analyzing the repeat purchase data and timeliness information in the purchased product feature information, the user's purchase preference characteristics are obtained.This can help the platform understand users' purchasing habits and changes in demand, and provide a basis for personalized recommendations and search result optimization. At the same time, purchasing behavior is a direct reflection of user needs. Repurchase data and timeliness information can reflect users' preference for products and the pattern of purchase time, which is of great significance to improving the quality of search services. The extended semantic features and the purchase preference features are then subjected to segmented product quantization compression to obtain compressed semantic association features, and the explicit semantic features and the compressed semantic association features are simultaneously subjected to double-layer compact indexing to obtain explicit results and associated results. The data dimension is reduced through segmented product quantization compression, which improves data processing efficiency. At the same time, double-layer compact indexing is performed to quickly obtain explicit results and associated results, making search results more accurate and efficient, and able to simultaneously provide potential user intentions, such as "breathable fabrics" and "vacation style". Finally, the private domain e-commerce search results are visualized based on the explicit results and the associated results, and then the explicit results and extended associated results are displayed to users in an intuitive manner, making it convenient for users to quickly find the required products and improve user experience and shopping efficiency. Visual presentation can allow users to more clearly understand the content and characteristics of search results, improve user satisfaction with the platform, and also avoid the problem of users having to repeat searches multiple times.

[0073] In one embodiment, the step S2 of acquiring style preference features based on the multiple click counts of the browsed products and the multiple click dwell times of the browsed products includes:

[0074] S201, sorting the number of clicks on a plurality of browsed products in descending order to obtain a ranking table of the number of clicks on browsed products;

[0075] S202, sorting the plurality of browsing click and dwell time in descending order to obtain a browsing click and dwell time sorting table;

[0076] S203, obtaining the first product corresponding to the first product with the highest number of clicks according to the browsed product click count ranking table;

[0077] S204: Obtain the second product corresponding to the first product in the first-place ranking list based on the browsing click and dwell time ranking list, apply dynamic load coefficient constraints to the second product and the first product to obtain a balanced cluster set, and perform centroid screening on the balanced cluster set based on the K-means clustering algorithm to obtain the product centroid;

[0078] S205: Screening the product centroid for segmented semantic features based on an attention mechanism to obtain multiple segmented semantic features;

[0079] S206, performing keyword extraction on the plurality of segmented semantic features based on the BERT language model to obtain a first keyword, and capturing context association of the first keyword based on a multi-head self-attention mechanism to obtain a keyword semantic text;

[0080] S207: Extracting style preference features of the keyword semantic text based on a preset text library.

[0081] As described in steps S201-S207 above, the present invention first sorts the multiple browsed products by click count in descending order to obtain a browsed product click count ranking table. Sorting the number of clicks by users on browsed products can intuitively demonstrate the differences in user interest in different products. Products with a high number of clicks are likely to be of greater interest to users. This ranking allows for rapid identification of these key products, providing a clear direction for subsequent analysis and helping e-commerce platforms accurately grasp user interest trends. The multiple browsed click dwell time is then sorted in descending order to obtain a browsed click dwell time ranking table. The browsed click dwell time reflects the depth of user interest in a product. The longer the dwell time, the greater the user's interest in the product's details, introduction, and other content, and the deeper their understanding of the product. This ranking further explores the user's potential interest in the product, assisting in determining the user's true preferences. This complements the click count ranking results to more comprehensively characterize user interests. Secondly, the first product corresponding to the top browsed product with the highest number of clicks is obtained from the browsed product click count ranking table. This obtains the top product from the ranking table, i.e., the product with the most clicks, identifying one of the products that the user is most interested in, and using it as a key sample for analyzing user style preferences. This product is representative, and its style characteristics are likely to be liked by users, providing important clues for subsequent style analysis. Next, the second product corresponding to the first product in the browsing click and dwell time ranking table is obtained according to the browsing click and dwell time ranking table, and the second product and the first product are subjected to dynamic load coefficient constraints to obtain a balanced cluster set. The balanced cluster set is subjected to centroid screening based on the K-means clustering algorithm to obtain the product centroid, and the second product with the longest dwell time is obtained. Combined with the first product, the products corresponding to the user's attention frequency and attention depth are comprehensively considered. Dynamic load factor constraints can balance the relationship between the two and prevent a single factor from dominating. The dynamic load factor constraint is a dynamic weight allocation mechanism. By weighting the contributions of different feature dimensions (such as click count and dwell time) in the clustering process, the resulting set of clusters can evenly reflect multi-dimensional user preferences and prevent a single feature from dominating the clustering results. The K-means clustering algorithm performs centroid screening, clustering products with similar characteristics into a single category and identifying the centroid that represents this category. This provides a basis for subsequent precise analysis of user style preferences, making the analysis results more universal and representative. The centroids of these products are then filtered for segmented semantic features based on the attention mechanism, resulting in multiple segmented semantic features. The attention mechanism is a deep learning technique that simulates the "selective attention" process in human cognition. Its core is to use dynamic weight allocation to enable the model to focus on the most relevant parts of the input information when processing the current task. This allows the attention mechanism to focus on the key semantic features of the product centroids and filter out irrelevant information.By filtering multiple segmented semantic features, a more detailed characterization of the product centroid is achieved, delving deeper into specific style elements that may be of interest to users. This provides more accurate information for subsequent keyword extraction and style preference determination. For example, in this solution, dress B ranks first in the click-to-dwell time ranking table, becoming the second item. Models A and B are dynamically constrained by load factors and then clustered using the K-means algorithm. If other dress styles are also included in the cluster, the resulting product centroid may exhibit a combination of features such as floral print, slim fit, and midi length. These features better represent Xiao Zhang's overall style preference for dresses. Keyword extraction is then performed on these segmented semantic features using the BERT language model to obtain the first keyword. The multi-head self-attention mechanism then contextualizes this first keyword to produce a keyword semantic text. The BERT language model's powerful semantic understanding allows accurate extraction of key information from the segmented semantic features, resulting in the first keyword. The multi-head self-attention mechanism captures the contextual associations of keywords in different contexts, enriching the semantic expression of keywords and producing a more comprehensive and contextualized keyword semantic text. This enables a deeper and more accurate understanding of user style preferences, providing strong support for determining the final style preferences. For example, based on segmented semantic features, the BERT language model extracts the first keywords such as "floral" and "slim fit." Through the multi-head self-attention mechanism, combined with the relevant context of the dress, keyword semantic text is obtained, such as "retro floral dress, slim design highlights body curves," which more comprehensively describes the dress style that the user is interested in. Finally, the style preference features of the keyword semantic text are extracted based on the preset text library, which stores definitions and descriptions of various styles. By matching with the keyword semantic text, the user's style preference features can be quickly determined. This converts complex semantic information into specific style categories, enabling e-commerce platforms to provide users with more accurate search results and personalized recommendations based on these features.

[0082] In one embodiment, the step S3 of performing semantic extraction on the search keyword to obtain explicit semantic features, and obtaining extended semantic features based on the explicit semantic features and style preference features includes:

[0083] S301, performing word segmentation processing on the search keyword to obtain multiple independent words of the search keyword, and converting the multiple independent words into word vectors based on the Word2Vec model to obtain independent word set vectors, and defining semantic features corresponding to the independent word set vectors as explicit semantic features;

[0084] S302, mapping the style preference feature into a style semantic vector aligned with the independent word set vector dimension through a fully connected layer;

[0085] S303, comparing and screening the independent word set vector and the style semantic vector based on a cosine similarity model to obtain an independent word-style semantic subset vector;

[0086] S304, concatenating the independent word-style semantic subset vector and the style semantic vector based on a preset time sequence to obtain a concatenated vector;

[0087] S305 , inputting the splicing vector into a preset bidirectional LSTM network for contextual semantic enhancement to obtain an enhanced splicing vector, and using the semantic features of the enhanced splicing vector as extended semantic features.

[0088] As described in the above steps S301-S305, the present invention first performs word segmentation processing on the search keyword to obtain multiple independent words of the search keyword, and converts the multiple independent words into word vectors based on the Word2Vec model to obtain independent word set vectors, and defines the semantic features corresponding to the independent word set vectors as explicit semantic features. The complex search keywords are split into basic units through word segmentation processing, which is convenient for model understanding and processing. The Word2Vec model converts words into vectors so that the semantics of the words can be presented in numerical form. Among them, Word2Vec is a word embedding technology based on neural networks, which is used to map words in natural language into low-dimensional dense vectors (usually 50 to 300 dimensions) so that words with similar semantics are close in distance in the vector space. The independent word set vector integrates the semantic information of the keyword, and the explicit semantic features clarify the direct semantics of the keyword, providing a basis for subsequent semantic expansion. This helps the model accurately grasp the core content of the user's search intent and provides a basis for the preliminary screening of search results. Secondly, the style preference features are mapped into style semantic vectors aligned with the dimensions of the independent word set vectors through the fully connected layer. The style preference features usually have a specific representation form, and the fully connected layer can transform them so that their dimensions are consistent with the independent word set vectors. In this way, the style preference features can be compared and integrated with the semantic vectors of the keywords in the same space, making it easier to explore the potential connection between the two, providing style information supplements for the extended semantic features, and making the search results more in line with the user's personalized style needs. For example: if the previous step analyzes that the user's style preference feature is "sports style with trendy elements", the fully connected layer is used to map it into a style semantic vector with the same dimension as the above-mentioned independent word set vector. For example, the style preference of "sports style with trendy elements" is encoded into a vector form so that it is consistent with the vector dimensions corresponding to "fashion" and "sneakers". In this way, in the subsequent steps, the style preference can be better combined with the semantics of the search keyword, and the independent word set vector and the style semantic vector are compared and screened based on the cosine similarity model to obtain the independent word-style semantic subset vector, and the cosine similarity model can measure the similarity between the two vectors. Through comparison and screening, independent words and their related semantic parts with high similarity to the style semantic vector can be found to form a subset vector. This step can filter out the parts related to the user's style preference from the keyword semantics, highlight the semantic information that matches the style, provide a more accurate direction for expanding the semantics, reduce the interference of irrelevant information, and improve the matching degree between the search results and the user's style preferences. For example: calculate the cosine similarity between the independent word set vector composed of "fashion" and "sports shoes" and the style semantic vector of "sports style with trendy elements".The "trend" component of "fashion" and the "sports" attribute of "sneakers" are found to have a high degree of similarity with the style semantic vector. This results in an independent word-style semantic subset vector, including semantic vectors related to "fashion (trend component)" and "sneakers (sports attribute)." During searches, the platform prioritizes products with these semantic characteristics, such as athletic-style sneakers with trendy design elements. The independent word-style semantic subset vector and the style semantic vector are then concatenated based on a preset time sequence to produce a concatenated vector. This pre-set time sequence provides an orderly method for vector concatenation, concatenating the subset vectors and style semantic vectors in a specific order, integrating semantic information about keywords and style preferences. The concatenated vector contains richer semantic content, encompassing both keyword- and style-related parts and overall style information. This provides more comprehensive data for subsequent semantic enhancement and helps generate extended semantic features that better meet user needs. Finally, the concatenated vector is input into a pre-set bidirectional LSTM network for contextual semantic enhancement, resulting in an enhanced concatenated vector. The semantic features of the enhanced concatenated vector are then used as extended semantic features. This allows the bidirectional LSTM network to simultaneously learn both forward and backward information from the input sequence, fully capturing the contextual semantics. After processing the concatenated vector, the resulting enhanced concatenated vector is semantically richer and more accurate. The extended semantic features generated based on this process can better reflect the underlying meaning of the user's search intent, ensuring that search results not only include directly relevant content but also contextually relevant information, improving search comprehensiveness and accuracy. For example, when the concatenated vector is input into the bidirectional LSTM network, the network learns the contextual association between "fashionable sneakers" and "sporty style with trendy elements," such as how trendy sneakers can be paired with sports scenes and which sporty and fashionable sneakers are suitable for different seasons. The resulting enhanced concatenated vector incorporates this rich contextual semantics, and its semantic features become extended semantic features. When searching, the platform may recommend trendy sneaker combinations suitable for different sports scenarios and popular sports-style sneaker styles of the season based on these extended semantic features, in addition to displaying ordinary fashionable sneakers.

[0089] In one embodiment, the step S4 of acquiring repeat purchase data information and timeliness information based on the purchased commodity characteristic information, and acquiring purchase preference characteristics based on the repeat purchase data information and timeliness information, includes:

[0090] S401: Obtain historical purchase timestamps and multiple product categories based on the purchased product feature information, and obtain the time interval between two adjacent purchases in each of the same product categories;

[0091] S402: Mark multiple behaviors whose time intervals are less than a preset threshold as repeat purchase events, and define data information corresponding to the repeat purchase events as repeat purchase data information;

[0092] S403: Obtain a time sliding window based on the historical purchase timestamps, and obtain a seasonal distribution interval based on the time sliding window, divide the multiple product categories into the seasonal distribution intervals, obtain seasonal interval naturalized product categories, and use time information of the seasonal interval naturalized product categories in the seasonal distribution intervals as timeliness information;

[0093] S404: Perform time-efficiency screening on the repurchase data information according to the time-efficiency information to obtain screened repurchase data information, and perform encoding processing on the screened repurchase data information to obtain a screened repurchase data code;

[0094] S405. Convert the filtered repurchase data encoding into a word vector based on the Word2Vec model to obtain the filtered repurchase data encoding vector, and use the filtered repurchase data encoding vector as a purchase preference feature.

[0095] As described in the above steps S401-S405, the present invention first obtains the historical purchase timestamps and multiple product categories based on the purchased product feature information, and obtains the time interval between two adjacent purchases under each of the same product categories. In this way, by obtaining the historical purchase timestamps and product categories, the time context of the user's purchase behavior and the types of goods involved can be clearly presented. Calculating the time interval between two adjacent purchases under the same product category can provide insights into the differences in the user's purchase frequency for different types of goods. This helps the e-commerce platform understand the user's consumption rhythm, which is crucial for judging the user's demand stability and repurchase potential for various types of goods. Then, multiple behaviors with a time interval less than a preset threshold are marked as repurchase events, and are defined as repurchase data information based on the data information corresponding to the repurchase event. Setting a preset threshold to define repurchase events can accurately screen out the user's repeated purchase behavior. Repurchase data information reflects the user's recognition and continuous demand for specific products, which is an important basis for judging user purchasing preferences. Based on this, the e-commerce platform can understand which products are popular with users, providing strong support for precision marketing and recommendations. Then, a time sliding window is obtained based on the historical purchase timestamps, and a seasonal distribution interval is obtained based on the time sliding window. The multiple product categories are divided into seasonal distribution intervals to obtain seasonal interval-normalized product categories. The time information of the seasonal interval-normalized product categories in the seasonal distribution interval is used as timeliness information. Through the use of time sliding windows and seasonal distribution intervals, the user's purchasing behavior is combined with time and seasonal factors. Among them, the time sliding window is a technical means widely used in time series data processing. It analyzes data by setting a fixed-length window on the time series and moving it step by step along the time axis. Its core principle is to use the movement of the window to dynamically observe and analyze local data features in the time series. By dividing the product categories by season, the seasonal pattern of user purchasing behavior can be discovered, that is, the difference in demand for different product categories in different seasons. Acquiring timeliness information provides e-commerce platforms with a more detailed understanding of user purchase time preferences, helping to recommend relevant products to users in appropriate seasons, improving the timeliness and accuracy of recommendations. For example, using user A's purchase data, using a time sliding window and seasonal segmentation reveals that she frequently purchases baby clothing in spring (March-May), primarily in categories such as light coats and onesies. These product categories are classified into the spring seasonal distribution interval and become seasonally normalized product categories. The spring time information represents the timeliness information associated with these products.In the spring, the platform can specifically recommend more clothing styles suitable for babies to user A. Then, the platform will perform time-sensitive screening on the repurchase data information according to the time-sensitive information to obtain the filtered repurchase data information. The filtered repurchase data information will be encoded to obtain the filtered repurchase data code. In this way, using time-sensitive information to filter the repurchase data information can further focus on the product data with repurchase behavior in a specific season and exclude interference data that does not meet seasonal needs. The encoding process converts the filtered repurchase data into a format suitable for model processing, which is convenient for subsequent in-depth analysis and feature extraction, so that the obtained purchase preference features more accurately reflect the actual purchase needs of users in different seasons. Secondly, based on the Word2Vec model, the filtered repurchase data code is converted into a word vector to obtain a filtered repurchase data code vector. The filtered repurchase data code vector is used as a purchase preference feature. The filtered repurchase data code is converted into a word vector through the Word2Vec model, so that the semantic information of the data can be represented in vector form, which is convenient for the model to calculate and analyze. The filtered repurchase data code vector integrates the user's repurchase product information in a specific season and comprehensively reflects the user's purchase preference characteristics. E-commerce platforms can use these vector features to more accurately provide users with product recommendations and optimize search results that match their purchasing preferences, improving both user satisfaction and platform operational efficiency. For example, consider the encoded data for user A's spring repeat purchases, converted into vectors using the Word2Vec model. Suppose the encoding vector for milk powder is represented as [0.1, 0.3, 0.5, ...], the encoding vector for diapers is [0.2, 0.4, 0.6, ...], and the encoding vector for baby coats is [0.7, 0.1, 0.8, ...]. These vectors are combined to represent user A's spring purchase preferences. When users search for related products, the platform prioritizes milk powder brands, diaper styles, and baby coat styles that match user A's preferences based on these vector features.

[0096] In one embodiment, the step S5 of performing piecewise product quantization compression on the extended semantic features and the purchase preference features to obtain compressed semantic association features includes:

[0097] S501, mapping the extended semantic feature into an extended semantic feature vector;

[0098] S502: Mapping the purchase preference feature into a purchase preference feature vector, and concatenating the extended semantic feature vector and the purchase preference feature vector to obtain a joint vector;

[0099] S503: Divide the joint vector into multiple sub-vector segments according to a preset length, and cluster each sub-vector segment based on K-means clustering to obtain multiple cluster centroids, where each cluster centroid represents an index;

[0100] S504: Map the plurality of cluster centroids into a plurality of cluster centroid vectors, and perform product quantization on the plurality of cluster centroid vectors to obtain particle vector values, and use the semantics corresponding to the particle vector values ​​as compressed semantic association features.

[0101] As described in the above steps S501-S504, in the field of deep learning and data processing, vector is a commonly used data representation form. Converting semantic features into vectors is in line with the way computers process data, and can make complex semantic information more efficiently transmitted and processed in the model. The present invention first maps the extended semantic features into extended semantic feature vectors, and converts the extended semantic features into vector form, which allows the computer to understand and process these semantic information in a numerical way. The vector form is convenient for various mathematical operations and analyses. After the semantic features are quantified, they can be compared, fused, and other operations with other vectors in subsequent steps, thereby digging out deeper semantic associations and improving the accuracy of understanding user search intentions. Secondly, the purchase preference features are mapped into purchase preference feature vectors, and the extended semantic feature vectors and the purchase preference feature vectors are spliced ​​to obtain a joint vector. The purchase preference features are also converted into vectors and spliced ​​with the extended semantic feature vectors, thereby realizing the fusion of information on user search intentions and purchasing habits. The joint vector contains more comprehensive information about user needs. It not only includes semantic extensions related to the current search keyword but also incorporates preferences reflected by the user's past purchase behavior. This provides a rich data foundation for more precise search result filtering and matching. The joint vector is then partitioned into multiple sub-vector segments of a preset length. Each sub-vector segment is clustered using K-means clustering to obtain multiple cluster centroids, each representing an index. This division of the joint vector into sub-vector segments reduces data dimensionality and computational complexity. K-means clustering groups similar sub-vector segments together and finds their cluster centroids. Each cluster centroid acts as an index, representing a class of vectors with similar characteristics. This facilitates classification and rapid retrieval of large amounts of data. This approach allows data with similar user needs to be grouped together, improving search efficiency and accuracy. For example, if the joint vector is 10 in length and is partitioned into sub-vector segments of length 3, the resulting sub-vector segments are 3 (the last segment with fewer than 3 elements can be padded or specially processed). Using the K-means clustering algorithm on these subvector segments, assuming K is 3 (dividing them into three categories), we obtain three cluster centroids: [0.7, 0.6, 0.8], [0.5, 0.4, 0.3], and [0.9, 0.8, 0.7]. These three cluster centroids represent different categories of user demand characteristics. For example, the first cluster centroid represents a category with high requirements for sports watch functionality and performance. Each cluster centroid is stored as an index in the database.When there is a new search request, the matching product data can be quickly found through these indexes. Finally, the multiple cluster centroids are mapped to multiple cluster centroid vectors, and the multiple cluster centroid vectors are multiplied and quantized to obtain particle vector values. The semantics corresponding to the particle vector values ​​are used as compressed semantic association features. In this way, the cluster centroids are mapped to vectors again and multiplied and quantized, which further compresses the data dimension and reduces storage space and calculation amount. The compressed semantic association feature retains the core semantic association information of the data, while ensuring the accuracy of the search, it improves the operating efficiency of the system. This compressed feature can be used more efficiently for subsequent index construction and search matching, making the search process faster and more accurate.

[0102] In one embodiment, the step S5 of simultaneously performing double-layer compact indexing on the explicit semantic features and the compressed semantic association features to obtain explicit results and association results includes:

[0103] S505: construct the explicit semantic feature based on the explicit feature inverted index to obtain a first-level index, and map the first-level index to a hash bucket based on locality sensitive hashing (LSH) to obtain a hash first-level index, and use the result corresponding to the hash first-level index as the explicit result;

[0104] S506: constructing a second-level index from the compressed semantic association features based on the compressed feature graph index, and constructing the second-level index into a multi-layer graph index based on a hierarchical navigable small-world (HNSW) graph algorithm, wherein each layer of the graph index corresponds to a product identifier;

[0105] S507: Use the result corresponding to the multi-layer graph index as the association result.

[0106] As described in steps S505-S507 above, the present invention first constructs the explicit semantic features based on an explicit feature inverted index to obtain a first-level index. This first-level index is then mapped to a hash bucket using locality-sensitive hashing (LSH) to obtain a hashed first-level index. The results corresponding to the hashed first-level index are then used as explicit results. This explicit feature inverted index can quickly locate documents (product data) containing specific explicit semantic features, improving retrieval efficiency. Locality-sensitive hashing (LSH) maps similar vectors to the same hash bucket, reducing the comparison scope and further accelerating retrieval. The explicit results provide users with basic search results directly related to the search keywords, meeting their basic needs and allowing them to quickly obtain the most direct and relevant product information. A second-level index is then constructed based on a compressed feature graph index using the compressed semantic association features. This second-level index is then constructed into a multi-layer graph index using a hierarchical navigable small-world (HNSW) graph algorithm. Each layer of the graph index corresponds to a product identifier. By constructing a compressed feature graph index based on the compressed semantic association features, this complex semantic information can be effectively organized and stored. The multi-layer graph index constructed by the HNSW graph algorithm enables efficient approximate nearest neighbor search. This method allows the system to quickly locate relevant product identifiers based on the user's potential needs (compressed semantic association features), identify products potentially related to the user's search intent, enrich search results, and improve search accuracy and comprehensiveness. Finally, the results corresponding to the multi-layer graph index are presented to the user as associated results, thus supplementing and expanding the search results. These associated results, based on the user's potential needs and behavior patterns, provide users with more products that may be of interest but have not yet been explicitly searched for, enhancing the user's shopping experience and increasing product exposure and sales potential.

[0107] This application also provides a private domain e-commerce data search system based on big data, including:

[0108] The first acquisition module is used to obtain user search feature information of the private domain e-commerce platform, wherein the user search feature information includes behavioral feature information and purchase feature information;

[0109] A second acquisition module is configured to acquire browsing characteristic information and search keywords based on the behavioral characteristic information, and acquire multiple browsed product click counts and multiple browsed click dwell times based on the browsing characteristic information, and acquire style preference characteristics based on the multiple browsed product click counts and multiple browsed click dwell times;

[0110] An extraction module, configured to perform semantic extraction on the search keyword to obtain explicit semantic features, and to obtain extended semantic features based on the explicit semantic features and style preference features;

[0111] A third acquisition module is configured to acquire repeat purchase data information and timeliness information based on the purchased commodity characteristic information, and acquire purchase preference characteristics based on the repeat purchase data information and timeliness information;

[0112] a fourth acquisition module, configured to perform piecewise product quantization compression on the extended semantic features and the purchase preference features to obtain compressed semantic association features, and to simultaneously perform double-layer compact indexing on the explicit semantic features and the compressed semantic association features to obtain explicit results and association results;

[0113] A presentation module is used to visually present the private domain e-commerce search results based on the explicit results and the associated results.

[0114] In one embodiment, the second acquisition module includes:

[0115] A first sorting unit is used to sort the click counts of the browsed products in descending order to obtain a sorting table of click counts of browsed products;

[0116] A second sorting unit is used to sort the plurality of browsing click and stay times in descending order to obtain a browsing click and stay time sorting table;

[0117] A first acquiring unit is configured to acquire, according to the browsed product click count ranking table, a first product corresponding to a first browsed product with a highest number of clicks;

[0118] A second acquisition unit is configured to acquire a second product corresponding to a first product in the browsing click and dwell time ranking table according to the browsing click and dwell time ranking table, and subject the second product and the first product to dynamic load coefficient constraints to obtain a balanced cluster set, and perform centroid screening on the balanced cluster set based on a K-means clustering algorithm to obtain a product centroid;

[0119] A screening unit, configured to screen the product centroid for segmented semantic features based on an attention mechanism to obtain multiple segmented semantic features;

[0120] A first extraction unit is configured to perform keyword extraction on the plurality of segmented semantic features based on a BERT language model to obtain a first keyword, and to capture contextual associations of the first keyword based on a multi-head self-attention mechanism to obtain a keyword semantic text;

[0121] The second extraction unit is configured to extract the style preference features of the keyword semantic text based on a preset text library.

[0122] The present application also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0123] The present application also provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when executed by a processor.

[0124] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct RAM bus dynamic RAM (DRDRAM), and RAM bus dynamic RAM (RDRAM).

[0125] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0126] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A private domain e-commerce data search method based on big data, characterized in that: include: Obtain user search feature information on the private domain e-commerce platform, where the user search feature information includes behavioral feature information and purchase feature information; Obtaining browsing characteristic information and search keywords based on the behavioral characteristic information, obtaining multiple browsed product click counts and multiple browse click dwell times based on the browsing characteristic information, and obtaining style preference characteristics based on the multiple browsed product click counts and multiple browse click dwell times; Performing semantic extraction on the search keyword to obtain explicit semantic features, and obtaining extended semantic features based on the explicit semantic features and style preference features; Obtaining repeat purchase data information and timeliness information based on the purchased product characteristic information, and obtaining purchase preference characteristics based on the repeat purchase data information and timeliness information; The extended semantic features and the purchase preference features are subjected to piecewise product quantization compression to obtain compressed semantic association features, and the explicit semantic features and the compressed semantic association features are subjected to simultaneous double-layer compact indexing to obtain explicit results and association results, wherein the specific implementation steps are as follows: The explicit semantic feature is constructed based on the explicit feature inverted index to obtain a first-level index, and the first-level index is mapped to a hash bucket based on locality sensitive hashing (LSH) to obtain a hash first-level index, and a result corresponding to the hash first-level index is used as the explicit result; constructing a second-layer index from the compressed semantic association features based on a compressed feature graph index, and constructing the second-layer index into a multi-layer graph index based on a hierarchical navigable small-world (HNSW) graph algorithm, wherein each layer of the graph index corresponds to a product identifier; The result corresponding to the multi-layer graph index is used as the association result; The private domain e-commerce search results are visualized based on the explicit results and the associated results.

2. The method for searching private domain e-commerce data based on big data according to claim 1, characterized in that: The step of obtaining style preference features based on the plurality of browsed product click times and the plurality of browsed click dwell times comprises: sorting the plurality of browsed products by the number of clicks in descending order to obtain a browsed product click number sorting table; Sort the plurality of browsing click and stay times in descending order to obtain a browsing click and stay time sorting table; Obtaining the first product corresponding to the first product click count according to the browsed product click count ranking table; The second product corresponding to the first product in the first-order browsing click and dwell time ranking table is obtained based on the browsing click and dwell time ranking table, and the second product and the first product are subjected to dynamic load coefficient constraints to obtain a balanced cluster set. The balanced cluster set is subjected to centroid screening based on the K-means clustering algorithm to obtain the product centroid. The dynamic load coefficient constraint is a dynamic weight distribution mechanism that adjusts the contribution weights of different feature dimensions in the clustering process by weighting, so that the finally generated cluster set can evenly reflect multi-dimensional user preferences and avoid a single feature dominating the clustering results. Performing segmented semantic feature screening on the product centroid based on an attention mechanism to obtain multiple segmented semantic features; Perform keyword extraction on the plurality of segmented semantic features based on the BERT language model to obtain a first keyword, and capture contextual associations of the first keyword based on a multi-head self-attention mechanism to obtain a keyword semantic text; The style preference features of the keyword semantic text are extracted based on a preset text library.

3. The method for searching private domain e-commerce data based on big data according to claim 1, characterized in that: The step of performing semantic extraction on the search keyword to obtain explicit semantic features, and obtaining extended semantic features based on the explicit semantic features and style preference features includes: Performing word segmentation processing on the search keyword to obtain multiple independent words of the search keyword, converting the multiple independent words into word vectors based on the Word2Vec model to obtain independent word set vectors, and defining the semantic features corresponding to the independent word set vectors as explicit semantic features; Mapping the style preference features into style semantic vectors aligned with the independent word set vector dimensions through a fully connected layer; Comparing and screening the independent word set vector with the style semantic vector based on a cosine similarity model to obtain an independent word-style semantic subset vector; Performing vector splicing on the independent word-style semantic subset vector and the style semantic vector based on a preset time sequence to obtain a spliced ​​vector; The splicing vector is input into a preset bidirectional LSTM network for contextual semantic enhancement to obtain an enhanced splicing vector, and the semantic features of the enhanced splicing vector are used as extended semantic features.

4. The method for searching private domain e-commerce data based on big data according to claim 1, characterized in that: The step of obtaining repeat purchase data information and timeliness information based on the purchased commodity characteristic information, and obtaining purchase preference characteristics based on the repeat purchase data information and timeliness information, includes: Obtaining historical purchase timestamps and multiple product categories based on the purchased product feature information, and obtaining the time interval between two adjacent purchases under each of the same product categories; Marking multiple behaviors with time intervals less than a preset threshold as repeat purchase events, and defining data information corresponding to the repeat purchase events as repeat purchase data information; Obtaining a time sliding window based on the historical purchase timestamps, and obtaining a seasonal distribution interval based on the time sliding window, dividing the plurality of product categories into the seasonal distribution intervals to obtain seasonal interval naturalized product categories, and using time information of the seasonal interval naturalized product categories in the seasonal distribution intervals as timeliness information; Performing time-efficiency screening on the repurchase data information according to the time-efficiency information to obtain screened repurchase data information, and encoding the screened repurchase data information to obtain a screened repurchase data code; Based on the Word2Vec model, the filtered repurchase data encoding is converted into a word vector to obtain the filtered repurchase data encoding vector, and the filtered repurchase data encoding vector is used as the purchase preference feature.

5. The method for searching private domain e-commerce data based on big data according to claim 1, characterized in that: The step of performing segmented product quantization compression on the extended semantic features and the purchase preference features to obtain compressed semantic association features includes: Mapping the extended semantic features into an extended semantic feature vector; Mapping the purchase preference feature into a purchase preference feature vector, and concatenating the extended semantic feature vector and the purchase preference feature vector to obtain a joint vector; Dividing the joint vector into a plurality of sub-vector segments according to a preset length, and clustering each sub-vector segment based on K-means clustering to obtain a plurality of cluster centroids, wherein each cluster centroid represents an index; The plurality of cluster centroids are mapped into the plurality of cluster centroid vectors, and the plurality of cluster centroid vectors are multiplied and quantized to obtain particle vector values, and the semantics corresponding to the particle vector values ​​are used as compressed semantic association features.

6. A private domain e-commerce data search system based on big data, used to execute the private domain e-commerce data search method based on big data according to any one of claims 1 to 5, characterized in that: include: The first acquisition module is used to obtain user search feature information of the private domain e-commerce platform, wherein the user search feature information includes behavioral feature information and purchase feature information; A second acquisition module is configured to acquire browsing characteristic information and search keywords based on the behavioral characteristic information, and to acquire a plurality of browsed product click counts and a plurality of browsed click dwell times based on the browsing characteristic information, and to acquire a style preference feature based on the plurality of browsed product click counts and the plurality of browsed click dwell times; An extraction module, configured to perform semantic extraction on the search keyword to obtain explicit semantic features, and to obtain extended semantic features based on the explicit semantic features and style preference features; A third acquisition module is configured to acquire repeat purchase data information and timeliness information based on the purchased commodity characteristic information, and acquire purchase preference characteristics based on the repeat purchase data information and timeliness information; a fourth acquisition module, configured to perform piecewise product quantization compression on the extended semantic features and the purchase preference features to obtain compressed semantic association features, and to simultaneously perform double-layer compact indexing on the explicit semantic features and the compressed semantic association features to obtain explicit results and association results; A presentation module is used to visually present the private domain e-commerce search results based on the explicit results and the associated results.

7. A private domain e-commerce data search system based on big data according to claim 6, characterized in that: The second acquisition module includes: A first sorting unit is used to sort the click counts of the browsed products in descending order to obtain a sorting table of click counts of browsed products; A second sorting unit is used to sort the plurality of browsing click and stay times in descending order to obtain a browsing click and stay time sorting table; A first acquiring unit is configured to acquire, according to the browsed product click count ranking table, a first product corresponding to a first browsed product with a highest number of clicks; A second acquisition unit is configured to obtain, based on the browse click dwell time ranking table, a second product corresponding to the first product in the browse click dwell time ranking table, and subject the second product and the first product to a dynamic load coefficient constraint to obtain a balanced cluster set, and subject the balanced cluster set to centroid screening based on a K-means clustering algorithm to obtain a product centroid, wherein the dynamic load coefficient constraint is a dynamic weight allocation mechanism that adjusts the contribution weights of different feature dimensions in the clustering process by weighting, so that the finally generated cluster set can evenly reflect multi-dimensional user preferences and avoid a single feature dominating the clustering results; A screening unit, configured to screen the product centroid for segmented semantic features based on an attention mechanism to obtain multiple segmented semantic features; A first extraction unit is configured to perform keyword extraction on the plurality of segmented semantic features based on a BERT language model to obtain a first keyword, and to capture contextual associations of the first keyword based on a multi-head self-attention mechanism to obtain a keyword semantic text; The second extraction unit is configured to extract the style preference features of the keyword semantic text based on a preset text library.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images