Item matching method and item matching device

Abstract

The present invention discloses an item matching method and an item matching device, and relates to the field of computer technology. A specific implementation of the method includes: generating a seed scene set based on user behavior data and item attribute data, and determining a scene vector of the seed scene set; obtaining a target item, and marking a target seed scene tag for the target item based on the scene vector of the seed scene set; and obtaining a target matching item that matches the target item based on the target seed scene tag. This implementation can analyze user behavior data and item attribute data, generate a seed scene set, and improve user experience and matching accuracy.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to an item matching method and an item matching device. Background Art

[0002] E-commerce platforms gradually optimize services such as item recommendations. After discovering that a user is interested in an item, they recommend candidate items that go with the item to the user to maximize the user experience. Therefore, item pairing is one of the important means of item recommendation. Existing item pairing technologies are mainly divided into three methods: 1. Manually organize and maintain item pairing lists; 2. Mining item pairing relationships based on association rule analysis; 3. Mining item pairing relationships based on machine learning and deep learning technologies.

[0003] In the process of realizing the present invention, the inventors found that there are at least the following problems in the prior art: 1. The method based on manually organizing and maintaining the item matching list requires huge manpower and material resources, and the number of matching combinations is limited, which cannot meet the growing user needs; 2. The method based on association rule analysis to mine the item matching relationship, the obtained item matching relationship has poor interpretability, which is easy to cause user complaints, and the number of obtained item matching relationships is small; 3. The method based on machine learning and deep learning technology to mine the item matching relationship has limited applicable scenarios, and the number of obtained item matching relationships is small. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides an item matching method and an item matching device, which can improve user experience and matching accuracy.

[0005] To achieve the above objective, according to a first aspect of an embodiment of the present invention, a method for matching items is provided.

[0006] An item matching method according to an embodiment of the present invention includes: generating a seed scene set based on user behavior data and item attribute data, and determining a scene vector of the seed scene set; obtaining a target item, and marking a target seed scene tag for the target item based on the scene vector of the seed scene set; and obtaining a target matching item that matches the target item based on the target seed scene tag.

[0007] Optionally, generating a set of seed scenarios based on user behavior data and item attribute data includes: acquiring user order data, user search data, user browsing data and item attribute data; mining item matching pairs corresponding to the user order data to generate a first seed scenario corresponding to the user order data; determining click data corresponding to the user search data to generate a second seed scenario corresponding to the user search data; fusing similar browsing sequences corresponding to the user browsing data to generate a third seed scenario corresponding to the user browsing data; generating a fourth seed scenario based on model matching and item main matching and according to the item attribute data; wherein the first seed scenario, the second seed scenario, the third seed scenario and the fourth seed scenario constitute the set of seed scenarios.

[0008] Optionally, mining the item collocation pairs corresponding to the user order data to generate the first seed scenario corresponding to the user order data includes: acquiring the order items corresponding to the user order data, combining the order items to generate candidate item collocation pairs; calculating the confidence and support of the candidate item collocation pairs based on association rules, and then selecting the item collocation pairs corresponding to the user order data from the candidate item collocation pairs in combination with a preset association threshold; promoting the selected item collocation pairs to core word collocation pairs, and then expanding the core word collocation pairs using an item library to obtain final item collocation pairs; and clustering and integrating the final item collocation pairs based on a clustering algorithm to generate the first seed scenario.

[0009] Optionally, determining the click data corresponding to the user search data and generating the second seed scenario corresponding to the user search data includes: screening the search terms corresponding to the user search data to obtain a candidate search term set; calculating the similarity between the candidate search terms in the candidate search term set based on the bag-of-words vector, and then determining a similar search term set in combination with a preset similarity threshold; obtaining click item data corresponding to the similar search term set; filtering the click item data, and generating the second seed scenario using the filtered click item data.

[0010] Optionally, the fusing of similar browsing sequences corresponding to the user browsing data to generate a third seed scenario corresponding to the user browsing data includes: obtaining a user browsing sequence corresponding to the user browsing data; filtering the user browsing sequence corresponding to the user browsing data based on a sequence length rule and a sequence category rule to obtain a first user browsing sequence; training a click-through rate prediction model using the first user browsing sequence, filtering the first user browsing sequence based on the click-through rate prediction model to obtain a second user browsing sequence; determining vector information corresponding to the second user browsing sequence, and then clustering and integrating the second user browsing sequence in combination with a clustering algorithm to obtain the similar browsing sequence; and generating the third seed scenario based on items corresponding to the similar browsing sequences.

[0011] Optionally, the fourth seed scenario is generated based on the model matching and the main and suitable matching of items according to the item attribute data, including: obtaining main category items and accessory category items according to the item category data, and then selecting items that match the main category items from the accessory category items according to the model attribute data of the main category items to generate a fifth seed scenario; obtaining the main core word corresponding to the item library, selecting the applicable item corresponding to the main core word according to the item title information in the item library, and then establishing a correspondence between the main item corresponding to the main core word and the applicable item to generate a sixth seed scenario; wherein the fifth seed scenario and the sixth seed scenario constitute the fourth seed scenario.

[0012] Optionally, determining the scene vector of the seed scene set includes: for each seed scene in the seed scene set, determining the scene vector of each seed scene according to the following method: performing word segmentation processing on the item title under each seed scene to obtain the scene segmentation of each seed scene; calculating the score of the scene segmentation, and then selecting the scene keyword of each seed scene from the scene segmentation according to the score of the scene segmentation according to preset keyword selection conditions; obtaining the keyword vector of the scene keyword from the word segmentation vector library, and calculating the global vector of each seed scene according to the keyword vector; obtaining the item vector of each seed scene and the global vector of each seed scene, and calculating the scene vector of each seed scene.

[0013] Optionally, the scene vector based on the seed scene set is used to mark the target seed scene label to which the target object belongs, including: obtaining the item vector of the target object from an item vector library, and calculating the vector distance between the item vector of the target object and the scene vector of each seed scene; based on a preset vector distance threshold, selecting the target seed scene corresponding to the target object from the seed scene set according to the vector distance; and marking the target seed scene as the target seed scene label to which the target object belongs.

[0014] Optionally, the method also includes: performing word segmentation processing on the item titles in the item library to obtain item title segmentations; based on an unsupervised word vector algorithm, performing vectorization processing on the item title segmentations to obtain the segmentation vectors of the item title segmentations to construct the segmentation vector library; for each item in the item library, obtaining the segmentation vector corresponding to each item, calculating the segmentation vector weight corresponding to each item, and calculating the item vector of each item according to the segmentation vector corresponding to each item and the segmentation vector weight corresponding to each item to construct the item vector library.

[0015] Optionally, acquiring a target matching item that matches the target item according to the target seed scene label includes: acquiring an item belonging to the target seed scene label from an item library, and determining that the acquired item is a first candidate matching item; selecting a second candidate matching item having the same attribute as the target item from the first candidate matching items according to a preset attribute filtering rule; and selecting the target matching item that has a matching relationship with the target item from the second candidate matching items based on a trained matching relationship recognition model.

[0016] Optionally, before selecting the target matching item that has a matching relationship with the target item from the second candidate matching items based on the trained matching relationship recognition model, the method further includes: acquiring user order data and re-behavior data corresponding to the user order data; training the matching relationship recognition model according to the user order data and the re-behavior data using a learning algorithm; and after selecting the target matching item that has a matching relationship with the target item from the second candidate matching items, the method further includes: setting an item sorting rule, and sorting the target matching items according to the item sorting rule.

[0017] To achieve the above objective, according to a second aspect of an embodiment of the present invention, a device for matching items is provided.

[0018] An item matching device according to an embodiment of the present invention includes: a generation module, which is used to generate a seed scene set based on user behavior data and item attribute data, and determine the scene vector of the seed scene set; a marking module, which is used to obtain a target item and mark the target item with a target seed scene label based on the scene vector of the seed scene set; and an acquisition module, which is used to obtain a target matching item that matches the target item based on the target seed scene label.

[0019] Optionally, the generation module is also used to: obtain user order data, user search data, user browsing data and item attribute data; mine item matching pairs corresponding to the user order data to generate a first seed scene corresponding to the user order data; determine the click data corresponding to the user search data to generate a second seed scene corresponding to the user search data; fuse similar browsing sequences corresponding to the user browsing data to generate a third seed scene corresponding to the user browsing data; generate a fourth seed scene based on model matching and item main matching and according to the item attribute data; wherein the first seed scene, the second seed scene, the third seed scene and the fourth seed scene constitute the seed scene set.

[0020] Optionally, the generation module is further used to: obtain order items corresponding to the user order data, combine the order items, and generate candidate item matching pairs; based on association rules, calculate the confidence and support of the candidate item matching pairs, and then select the item matching pairs corresponding to the user order data from the candidate item matching pairs in combination with a preset association threshold; promote the selected item matching pairs to core word matching pairs, and then use the item library to expand the core word matching pairs to obtain final item matching pairs; cluster and integrate the final item matching pairs based on a clustering algorithm to generate the first seed scenario.

[0021] Optionally, the generation module is also used to: filter the search terms corresponding to the user search data to obtain a candidate search term set; calculate the similarity between the candidate search terms in the candidate search term set based on the bag-of-words vector, and then determine a similar search term set in combination with a preset similarity threshold; obtain the click item data corresponding to the similar search term set; filter the click item data, and use the filtered click item data to generate the second seed scene.

[0022] Optionally, the generation module is also used to: obtain a user browsing sequence corresponding to the user browsing data; filter the user browsing sequence corresponding to the user browsing data based on a sequence length rule and a sequence category rule to obtain a first user browsing sequence; use the first user browsing sequence to train a click-through rate prediction model, filter the first user browsing sequence based on the click-through rate prediction model to obtain a second user browsing sequence; determine the vector information corresponding to the second user browsing sequence, and then cluster and integrate the second user browsing sequence in combination with a clustering algorithm to obtain the similar browsing sequence; generate the third seed scene according to the items corresponding to the similar browsing sequence.

[0023] Optionally, the generation module is also used to: obtain main category items and accessory category items according to item category data, and then select items that match the main category items from the accessory category items according to the model attribute data of the main category items to generate a fifth seed scene; obtain the main core word corresponding to the item library, select the applicable item corresponding to the main core word according to the item title information in the item library, and then establish a correspondence between the main item corresponding to the main core word and the applicable item to generate a sixth seed scene; wherein the fifth seed scene and the sixth seed scene constitute the fourth seed scene.

[0024] Optionally, the generation module is also used to: for each seed scene in the seed scene set, determine the scene vector of each seed scene according to the following method: perform word segmentation processing on the item title under each seed scene to obtain the scene segmentation of each seed scene; calculate the score of the scene segmentation, and then select the scene keyword of each seed scene from the scene segmentation according to the score of the scene segmentation according to preset keyword selection conditions; obtain the keyword vector of the scene keyword from the word segmentation vector library, and calculate the global vector of each seed scene according to the keyword vector; obtain the item vector of each seed scene and the global vector of each seed scene, and calculate the scene vector of each seed scene.

[0025] Optionally, the marking module is also used to: obtain the item vector of the target item from the item vector library, and calculate the vector distance between the item vector of the target item and the scene vector of each seed scene; based on a preset vector distance threshold, select the target seed scene corresponding to the target item from the seed scene set according to the vector distance; and mark the target seed scene as the target seed scene label to which the target item belongs.

[0026] Optionally, the device also includes a construction module, which is used to: perform word segmentation processing on the item titles in the item library to obtain item title segmentations; based on an unsupervised word vector algorithm, vectorize the item title segmentations to obtain the segmentation vectors of the item title segmentations to construct the segmentation vector library; for each item in the item library, obtain the segmentation vector corresponding to each item, calculate the segmentation vector weight corresponding to each item, and calculate the item vector of each item according to the segmentation vector corresponding to each item and the segmentation vector weight corresponding to each item to construct the item vector library.

[0027] Optionally, the acquisition module is further used to: acquire items belonging to the target seed scene label from an item library, and determine that the acquired items are first candidate matching items; select second candidate matching items having the same attributes as the target items from the first candidate matching items according to preset attribute filtering rules; and select the target matching items having a matching relationship with the target item from the second candidate matching items based on a trained matching relationship recognition model.

[0028] Optionally, the acquisition module is also used to: acquire user order data and repeated behavior data corresponding to the user order data; use a learning algorithm to train a matching relationship recognition model based on the user order data and the repeated behavior data; and set item sorting rules to sort the target matching items according to the item sorting rules.

[0029] To achieve the above objective, according to a third aspect of an embodiment of the present invention, an electronic device is provided.

[0030] An electronic device according to an embodiment of the present invention includes: one or more processors; a storage device for storing one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the item matching method according to the embodiment of the present invention.

[0031] To achieve the above objective, according to a fourth aspect of an embodiment of the present invention, a computer-readable medium is provided.

[0032] A computer-readable medium according to an embodiment of the present invention stores a computer program, and when the program is executed by a processor, the item matching method according to the embodiment of the present invention is implemented.

[0033] An embodiment of the above invention has the following advantages or beneficial effects: it is able to analyze user behavior data and item attribute data to generate a seed scene set, so as to cover as many item combinations as possible under the seed scene, solving the problem that the prior art cannot fully utilize item content information and user behavior information, and also solves the problem of a small number of item combinations and insufficient coverage of categories, thereby improving user experience and matching accuracy. In addition, in the embodiment of the present invention, a scene vector can also be determined, which is equivalent to extracting keywords for each item combination as the name of the item combination, thereby increasing the interpretability of the item combination and solving the problem of the interpretability of the item combination, which is of great significance for the actual implementation of the e-commerce platform.

[0034] The further effects of the above-mentioned non-conventional optional manner will be described below in conjunction with the specific implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings are used to better understand the present invention and do not constitute an improper limitation of the present invention.

[0036] Figure 1 is a schematic diagram of the main steps of the item matching method according to an embodiment of the present invention;

[0037] Figure 2 is a schematic diagram of generating a first seed scene by clustering and integrating the final item pairings based on a clustering algorithm according to an embodiment of the present invention;

[0038] Figure 3 is a schematic diagram of obtaining a similar search word set according to an embodiment of the present invention;

[0039] Figure 4 is a schematic diagram of main steps of a method for generating a seed scene set according to user behavior data and item attribute data according to an embodiment of the present invention;

[0040] Figure 5 is a schematic diagram of the main process of a method for determining a scene vector of a seed scene according to an embodiment of the present invention;

[0041] Figure 6 is a schematic diagram of the main process of a method for obtaining a target matching item that matches a target item according to an embodiment of the present invention;

[0042] Figure 7 is a schematic diagram of main modules of an item matching device according to an embodiment of the present invention;

[0043] Figure 8 is an exemplary system architecture diagram to which embodiments of the present invention may be applied;

[0044] Fig. 9 It is a schematic diagram of the structure of a computer system of a terminal device or a server suitable for implementing an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following is a description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0046] The main application scenario of the present invention is the collocation of items in the field of e-commerce. In the field of e-commerce, in order to improve the click-through rate and click-through conversion rate of items and improve the user experience, the e-commerce platform usually optimizes the item recommendation results, wherein the item collocation combination is one of the important means of item recommendation. In the early stage of e-commerce development, a large number of item collocation relationships were manually sorted and maintained based on the platform's own historical records and professional field knowledge. Although this method can provide a relatively accurate item collocation combination, it requires huge manpower and material resources, and the item collocation combinations provided are limited. In order to provide more accurate and more numerous item collocation relationships, e-commerce platforms have begun to explore how to automatically and accurately mine as many item collocation combinations as possible. Among them, the most common way is to mine the item collocation combination based on the association rule analysis technology, and on this basis, add other correlation metrics to improve the accuracy. There are also many challenges encountered in this method. On the one hand, the obtained item collocation combination has poor interpretability, and on the other hand, it is only applicable to user order records, and cannot be applied to other user behaviors, including but not limited to browsing, adding purchases, searching, etc., and at the same time, it also ignores the content information provided by the platform for the item itself, resulting in a limited combination of item collocations. With the development of machine learning and deep learning, e-commerce platforms have begun to explore methods based on machine learning and deep learning to identify item collocation relationships, and then obtain item collocation combinations. Different from association rule analysis technology, methods based on machine learning and deep learning pay more attention to the content information provided by the item itself, including item title, attributes, and images, but ignore user behavior information, resulting in poor recognition accuracy. At the same time, the user behavior information is not fully utilized, which also leads to a limited number of recognized item combinations.

[0047] In order to solve the above problems, the embodiments of the present invention propose an item matching method and an item matching device in an e-commerce scenario, which can mine possible item matching combinations based on user behavior data and item attribute data, thereby solving the problem that item attribute data and user behavior data are not fully utilized in the prior art, and also solving the problem that the prior art has a small number of item matching combinations and insufficient coverage of categories. Figure 1 FIG. 1 is a schematic diagram of the main steps of the method for matching items according to an embodiment of the present invention. Figure 1 As shown, the main steps of the item matching method may include:

[0048] Step S101, generating a seed scene set according to user behavior data and item attribute data, and determining a scene vector of the seed scene set;

[0049] Step S102, obtaining a target object, and marking a target seed scene label for the target object based on a scene vector of a seed scene set;

[0050] Step S103, obtaining a target matching item that matches the target item according to the target seed scene tag.

[0051] Among them, the seed scene set is composed of seed scenes. The seed scene can be defined as a combination of two or more types of items. Any two items in the combination have a matching relationship. The matching relationship can be defined as the existence of a certain type of people who may use them at the same time at a specific time or place. For example, fish hooks and baits, tents and picnic mats, and swimsuits and goggles are in a matching relationship. User behavior data refers to the behavior data of users on the e-commerce platform, which can include: user order data, user browsing data and user search data. User order data can be regarded as order data for users to purchase items or order data for users to add items to shopping. User browsing data refers to data for users to browse items. User search data refers to data for users to search for items on the platform. In the embodiment of the present invention, user order data, user browsing data and user search data can be analyzed to generate seed scenes related to user behavior, that is, corresponding item matching combinations are generated according to user behavior data, which solves the problem that the number of item matching combinations is small due to the failure to fully utilize user behavior data in the prior art. Item attribute data refers to attribute information data of items, such as item title information data, item model data, etc. In an embodiment of the present invention, the attribute data of the items can be analyzed to generate seed scenarios related to the item attributes, that is, corresponding item collocation combinations are generated according to the item attribute data, which solves the problems of limited applicable scenarios and poor interpretability of the obtained item collocation relationships in the prior art.

[0052] In addition, after generating the seed scene set, the scene vector of the seed scene set can be determined, that is, the scene vector of each seed scene in the seed scene set can be determined. The scene vector is equivalent to the keyword of the seed scene, that is, in the embodiment of the present invention, a keyword can be set for each item combination, and the set keyword can be used as the name of the item combination, thereby increasing the interpretability of the item combination, which is of great significance for the actual implementation of the e-commerce platform.

[0053] In step S101, a seed scene set is generated based on user behavior data and item attribute data, respectively. However, due to the increasing number of items on the e-commerce platform and the limited number of users on the e-commerce platform, the obtained seed scene cannot cover all items. In order to solve this phenomenon, after obtaining the target item, the embodiment of the present invention can mark the target item with the target seed scene label to which it belongs, so as to realize the classification of any item into the existing seed scene, greatly improving the coverage rate of the seed scene. Among them, the target item can be regarded as an item that needs to be obtained with a matching relationship, that is, the purpose of the embodiment of the present invention is to obtain an item that can be matched with the target item, the target seed scene is the seed scene corresponding to the selected target item, and the target seed scene belongs to the scene in the seed scene set. The target item can be an item from the user behavior data, for example, the target item can be a book browsed by the user, and the purpose is to recommend books that can be matched with the book to the user; the target item can also be an item that needs to be matched when setting an item package combination on the e-commerce platform, for example, the platform sets a mobile phone and a mobile phone case as an item package combination, and the target item can be regarded as a certain model of mobile phone, and the purpose is to match a mobile phone case suitable for the model of mobile phone.

[0054] The target matching item refers to an item that can be matched with the target item, and the target item and the target matching item form an item matching relationship. Through step S102, the target item can be marked with the target seed scene label to which it belongs, that is, the target seed scene can be determined. Therefore, in step S103, the item corresponding to the target seed scene can be obtained, so that the target matching item that can be matched with the target item can be selected from the obtained items.

[0055] The technical solution for item matching in the embodiment of the present invention can analyze user behavior data and item attribute data to generate a set of seed scenarios, thereby covering as many item matching combinations as possible in the seed scenarios, solving the problem that the prior art cannot fully utilize item content information and user behavior information, and also solves the problem of a small number of item matching combinations and insufficient coverage of categories, thereby improving user experience and matching accuracy. In addition, in the embodiment of the present invention, a scene vector can also be determined, which is equivalent to extracting keywords for each item matching combination as the name of the item matching combination, thereby increasing the interpretability of item matching and solving the problem of interpretability of item matching combinations, which is of great significance for the actual implementation of e-commerce platforms.

[0056] The generation of a seed scene set is an important part of the embodiment of the present invention. In the embodiment of the present invention, corresponding seed scenes can be generated based on user behavior data and item attribute data. Among them, user behavior data can include user order data, user search data, and user browsing data. For different types of data, corresponding seed scenes are generated, and then the generated seed scenes are used to form a seed scene set. As a reference embodiment of the present invention, the generation of a seed scene set based on user behavior data and item attribute data can include:

[0057] Step S1011, obtaining user order data, user search data, user browsing data and item attribute data;

[0058] Step S1012, mining item matching pairs corresponding to the user order data, and generating a first seed scene corresponding to the user order data;

[0059] Step S1013, determining click data corresponding to the user search data, and generating a second seed scenario corresponding to the user search data;

[0060] Step S1014, fusing similar browsing sequences corresponding to the user browsing data to generate a third sub-scenario corresponding to the user browsing data;

[0061] Step S1015: Generate a fourth sub-scenario based on the model matching and the item main matching according to the item attribute data.

[0062] In step S1011, user order data, user browsing data, user search data and item attribute data are obtained, which has been explained in detail in step S101 above and will not be repeated here. Steps S1012 to S1014 are used to analyze user behavior data and explore potential item matching combinations, that is, to generate seed scenes corresponding to different user behavior data, so as to provide scene labels for target items in the future. It can be seen that in the embodiment of the present invention, three different seed scene generation methods are proposed for the user's order, search and browsing data respectively.

[0063] (I) User order data

[0064] In a reference embodiment of the present invention, mining item matching pairs corresponding to user order data and generating a first seed scenario corresponding to the user order data may include steps S10121 to S10124.

[0065] Step S10121: Obtain order items corresponding to the user order data, combine the order items, and generate candidate item matching pairs.

[0066] It should be noted that the order items in step S10121 are items obtained after cleaning the user order data. Considering the actual situation, in order to increase the probability of user purchase or additional purchase, the platform usually subsidizes users through gift giving, free shipping, etc. These items will interfere with the subsequent item matching relationship. Therefore, in order to keep the user order data more in line with common sense, it is first necessary to filter based on the two attributes of "whether it is a gift" and "whether it is a special item" in the attributes of the item to obtain the cleaned user order data. "Whether it is a gift" and "whether it is a special item" can be identified by the platform itself; a gift refers to another item that may be unrelated to the item given away for free, such as a mobile phone sticker given on the platform, and the mobile phone sticker here is a gift; special items can be services that cannot be traded, such as shipping, warranty, etc.

[0067] After cleaning and filtering the user order data, the order items corresponding to the user order data can be obtained, and the order items can be combined to generate candidate item matching pairs. Specifically, all items under the same order are obtained and integrated into candidate item matching pairs by combining them two by two. This step is based on the assumption that items under the same order are usually more likely to match each other. For example, the user purchased three items as shown in Table 1 under the same order. Then, according to the order, the candidate item matching pairs obtained are: (5615776, 3369356), (5615776, 2856105), (3369356, 2856105).

[0068]

[0069] Table 1

[0070] Step S10122: Based on the association rules, the confidence and support of the candidate item pair are calculated, and then the item pair corresponding to the user order data is selected from the candidate item pair in combination with the preset association threshold.

[0071] The purpose of this step is to calculate the confidence and support between candidate item pairs based on association rule analysis, and then filter through empirical thresholds to obtain the final item pairs that meet the requirements. Among them, association rules are to statistically discover potential connections between data; confidence is also called reliability, or confidence level, confidence coefficient, that is, when sampling to estimate the overall parameter, due to the randomness of the sample, the conclusion is always uncertain. Therefore, a probability statement method is adopted, that is, the interval estimation method in mathematical statistics, that is, the estimated value and the overall parameter are within a certain allowable error range. The corresponding probability is called confidence; support is the degree of support, which indicates the frequency of the antecedent and the subsequent items appearing simultaneously in a data set. In the embodiment of the present invention, confidence (X→Y) refers to the probability that item X appears in the order and item Y appears at the same time. The calculation formula is as follows:

[0072]

[0073] Support refers to the probability that item X and item Y appear in an order at the same time. The calculation formula is as follows:

[0074]

[0075] Through the above two formulas, the corresponding confidence and support are calculated for each pair of candidate item pairings obtained in step S10121, and then the item pairings with low support and low confidence are filtered out through the empirical threshold to obtain the item pairings.

[0076] Step S10123: The selected item collocation pairs are promoted to core word collocation pairs, and then the core word collocation pairs are expanded using the item library to obtain final item collocation pairs.

[0077] The main purpose of this step is to increase the number of item collocation pairs. In step S10122, item collocation pairs that meet the confidence and support can be obtained. For example, the item collocation pairs that meet the requirements obtained from step S10122 are: (5615776, 3369356), (5615776, 2856105). According to the mapping table between items and core words, the core words of the items are extracted, as shown in Table 2. Among them, the mapping table between items and core words can be used to identify the core words of any item, and identifying the core words of an item belongs to the existing mature technology, and no specific explanation is provided in the embodiments of the present invention.

[0078]

[0079] Table 2

[0080] From Table 2, we can get that the core word of item number 5615776 is whipped cream, the core word of item number 3369356 is egg beater, and the core word of item number 2856105 is scraper. Then, in this step, we can get the core word granularity core word collocation pairs: (whipped cream, egg beater), (whipped cream, scraper). Then, map from the core word granularity to the item granularity to get the final item collocation pair. For the item library of the e-commerce platform, the core word of any item is obtained based on the mapping table of items and core words, and then the collocation pairs at the core word level are obtained, and all items belonging to the core word collocation pairs are integrated into the item collocation pairs, so as to obtain item collocation pairs with more data. For example, the core word collocation pair (whipped cream, egg beater) is obtained, then all items with core words of whipped cream and egg beater are taken out from the item library of the e-commerce platform, and are combined into item collocation pairs in the form of (whipped cream, egg beater), so as to obtain more item collocation pairs. The advantage of this step is that as many item collocation pairs as possible can be obtained on the basis of step S10122.

[0081] Step S10124: Clustering and integrating the final item pairings based on a clustering algorithm to generate a first seed scene.

[0082] The purpose of this step is to integrate all the acquired item collocation pairs and construct seed scenes. In this step, a graph is established with each item collocation pair as an edge, where nodes are items and edges are collocation relationships. Then, a clustering algorithm is used to obtain multiple combinations with collocation relationships, thereby obtaining multiple seed scenes. The clustering algorithm includes but is not limited to the connected subgraph clustering algorithm, the K-Means clustering algorithm, and the like. Figure 2 is a schematic diagram of generating a first seed scene by clustering and integrating the final item pair based on a clustering algorithm according to an embodiment of the present invention, such as Figure 2 As shown, each node is an item, and the text format of the node is "item number: core word". Each edge indicates that there is a collocation relationship between two nodes. In this step, since the four item nodes in the figure have a collocation relationship, we can get (5615776: whipped cream, 3369356: egg beater, 2856105: scraper, 12045358561: piping bag) as a seed scene.

[0083] (II) Searching for user data

[0084] In a reference embodiment of the present invention, determining click data corresponding to user search data and generating a second seed scenario corresponding to the user search data may include steps S10131 to S10134.

[0085] Step S10131: Filter the search terms corresponding to the user search data to obtain a set of candidate search terms.

[0086] The main purpose of this step is to find search terms that may have item pairing relationships. Common search terms can be roughly divided into two categories: 1. Search terms with clear intentions, such as a certain brand of mobile phone, washing machine, etc.; 2. Search terms with unclear intentions, such as outdoor camping, birthday gifts, etc. Among them, search terms with unclear intentions contain a large number of search terms with potential item pairing relationships, but this type of search terms also has some search terms with too vague intentions, such as platform self-operated, free shipping for orders over 100, one-yuan flash sales, etc. The main purpose of this step is to find out the search terms with unclear intentions. The main methods are as follows: (1) Identify search terms with clear intentions. Obtain all candidate search terms from the user's search logs for a period of time (which can be but is not limited to 2 months). Then use the mapping table between items and core words to obtain the core words of any item in the item library. Extract all the core words in the item library and organize them into a core word list. Then use the candidate search terms to match the core word list. If the candidate search terms contain words in the core word list, then the search term is considered to be a search term with clear intentions and can be filtered. (2) Identify search terms with overly vague intentions. This step requires manually organizing a list of meaningless search terms, such as free shipping, 10-yuan flash sales, platform self-operated, overseas purchases, etc., and match the list with the candidate search terms. If the match is successful, the search term is considered to be a search term with overly vague intentions and then filtered. After steps (1) and (2), the final set of candidate search terms is obtained.

[0087] Step S10132: Based on the bag-of-words vector, calculate the similarity between the candidate search words in the candidate search word set, and then determine the similar search word set in combination with a preset similarity threshold.

[0088] Since the same search intent can be expressed by different phrases, there are multiple search terms that express only one intent. The purpose of this step is to merge search terms with similar search intents and consider all clicks under search terms with similar search intents as common clicks. The specific method is: use common word segmentation tools to segment search terms, map search terms to vectors using bag-of-words vectors, and then calculate the vector distance between any two search term vectors, obtain similar search terms using empirical thresholds, and merge them. Among them, word segmentation tools include but are not limited to jieba word segmentation, Harbin Institute of Technology LTP, etc. For example, there are candidate search terms "XX squirrel snacks", "XX squirrel", and "XX squirrel flagship store". Their word segmentation results are: ("XX squirrel", "snacks"), ("XX squirrel"), ("XX squirrel", "flagship store"), and all the word segments are integrated into a word bag based on the word segmentation results: ("XX squirrel", "snacks", "flagship store"), then the word bag vectors corresponding to the three search terms are: [1,1,0], [1,0,0], [1,0,1]. The distance between the vectors is calculated to determine whether the search terms are similar. Here, the Euclidean distance is taken as an example: the Euclidean distance between "XX squirrel snacks" and "XX squirrel" is 1, and the two search terms are judged to be similar based on the empirical threshold.

[0089] Step S10133: Obtain clicked item data corresponding to a set of similar search terms.

[0090] In step S10132, it is determined whether any two candidate search terms have a similar relationship, thereby obtaining a relationship graph with search terms as nodes and similarity as edges, and finding a set of similar search terms based on a common connected subgraph clustering algorithm. According to user search data, the clicked item data under each candidate search term can be obtained. In this step, the clicked item data corresponding to each search term in the similar search term set can be considered as the clicked item data under the similar search term set. Figure 3 is a schematic diagram of obtaining a similar search word set according to an embodiment of the present invention, such as Figure 3 As shown in the figure, the nodes in the figure are search terms and the clicked item sequences under the search terms, and the edges represent the similarity relationship between the nodes. The three nodes in the figure have a similar relationship and can be considered as different search terms with the same search intent. Therefore, through the connected subgraph clustering method, it is obtained that the three nodes in the figure belong to the same similar search term set, and the click sequences under this set can be considered to belong to the same search intent, that is: the item number sequence [2,3,4,7,8] is the clicked item data under the similar search term set.

[0091] Step S10134: Filter the clicked object data, and use the filtered clicked object data to generate a second seed scene.

[0092] According to the definition of seed scenes, the number of candidate items in each seed scene must not be less than two. In order to prevent the phenomenon that there are many clicked item data under some search terms, but they are basically the same item, this step maps the items clicked under each search term to core words based on the mapping table of items and core words, and then removes duplicates and counts them, and filters the search terms with a core word sequence length of less than 2, so as to obtain the seed scene sequence under the search term. For example, the clicked item sequence corresponding to the search term "a certain model of mobile phone" is: [3133857, 10673889493, 57004994625], and the core word sequence corresponding to this item sequence is: [mobile phone, mobile phone, mobile phone]. After deduplication and counting, the search term "a certain model of mobile phone" is filtered out, and the clicked item data corresponding to this search term is filtered out, and then the filtered clicked item data is used to generate the second seed scene.

[0093] (III) User browsing data

[0094] In a reference embodiment of the present invention, similar browsing sequences corresponding to user browsing data are integrated to generate a third seed scenario corresponding to the user browsing data, which may include steps S10141 to S10145.

[0095] Step S10141: Obtain the user browsing sequence corresponding to the user browsing data.

[0096] In an embodiment of the present invention, a preset session time (which may be but is not limited to 3 hours) may be set, and the user browsing data within the preset session time is taken as one session, and the user's recent (which may be but is not limited to 1 month) browsing data is segmented according to time intervals, thereby generating a number of browsing session sequences for each user. For example, the browsing sequence of user A is: [sku1: 2019-11-12 17:33, sku2: 2019-11-12 18:20, sku3: 2019-11-12 19:56, sku4: 2019-11-12 21:00, sku5: 2019-11-12 21:59], and the preset session time is 3 hours, then the following browsing session sequence can be obtained: [sku1, sku2, sku3], [sku2, sku3, sku4], [sku3, sku4, sku5], [sku4, sku5], [sku5]. It should be noted that each browsed item corresponds to a session. The time range of the session is the sequence of browsed items from the time the item was browsed to the preset session time range. For example, the preset session time is 3 hours. Suppose browsing of item A starts at 13:00, then the items browsed in the next 3 hours should be placed in a sequence with A as the first element. Browsing of item B starts at 13:30, then all items in the next 3 hours should be placed in a sequence with B as the first element.

[0097] Step S10142: filtering the user browsing sequence corresponding to the user browsing data based on the sequence length rule and the sequence category rule to obtain a first user browsing sequence.

[0098] As described in step S10141, each user has several browsing session sequences, and the duration of each session sequence is the preset session time. Since the number of users on the e-commerce platform is huge and the click data is relatively messy, it is necessary to filter out the low-quality user session sequences. In this step, the dimensions of user browsing sequence filtering can be: (1) browsing sequence length filtering, selecting the upper and lower limits of the user browsing sequence length according to the experience value; (2) sequence length filtering of the category granularity in the browsing sequence, based on the existing browsing item sequence, converting the sequence into the first-level category granularity to ensure that all items in the user browsing item sequence are from the same first-level category, wherein the first-level category is the category information to which the item belongs, which refers to the item category pre-set by the e-commerce platform, for example, the item number is 3133857, the title is "a certain brand of mobile phone, 128G, black, China Mobile, China Unicom, China Telecom 4G mobile phone", and the first-level category to which it belongs is "mobile phone communication". In the embodiment of the present invention, the filtered sequence is named the first user browsing sequence.

[0099] Step S10143: using the first user browsing sequence to train a click rate prediction model, filtering the first user browsing sequence based on the click rate prediction model to obtain a second user browsing sequence.

[0100] The main purpose of this step is to train a click-through rate prediction model based on the first user browsing sequence, thereby achieving modeling of the user browsing sequence. The main purpose of the model is to estimate the probability of a user clicking on the next item when it is known that the user is browsing multiple items. The model training data comes from the first user browsing sequence, and the first n-1 items in each sequence are used as the user's browsing sequence to predict the user's click probability on the nth item. Among them, the architecture of the model can be a common deep learning model, including but not limited to a video recommendation model, a news recommendation model, etc. Then, based on the first user browsing sequence, the obtained click-through rate prediction model is used to predict any user sequence of each user, and the user browsing sequence with an estimated click-through rate greater than the empirical threshold is selected as the user's high-quality browsing sequence. In the embodiment of the present invention, the sequence obtained by filtering the model is named the second user browsing sequence.

[0101] Step S10144: determine the vector information corresponding to the browsing sequence of the second user, and then cluster and integrate the browsing sequence of the second user in combination with a clustering algorithm to obtain similar browsing sequences.

[0102] Based on the click rate prediction model obtained in step S10143, the click rate of the first user browsing sequence is predicted, and the vector information of the fully connected layer in the model structure can be extracted as the vector information of the user browsing sequence. Then, a clustering algorithm is used to obtain a set of similar user browsing sequences whose vector distances meet the empirical threshold, wherein the clustering algorithm includes but is not limited to the KMeans clustering algorithm, the local sensitive hashing algorithm, the Annoy algorithm, etc.

[0103] Step S10145: Generate a third sub-scene according to the items corresponding to the similar browsing sequence.

[0104] In step S10144, a similar browsing sequence can be obtained. Based on the similar browsing sequence, all items in the browsing sequence are merged so that each similar browsing sequence corresponds to a group of item sequences, thereby obtaining a final seed scene, wherein each similar browsing sequence is considered to be an independent seed scene.

[0105] Steps S1012 to S1014 are used to analyze user behavior data and mine potential item matching combinations, that is, to generate seed scenes corresponding to different user behavior data, so as to provide scene labels for target items in the future. Step S1015 is used to generate a fourth seed scene based on model matching and item main matching according to item attribute data. The main purpose of step S1015 is to make full use of item attribute data, including but not limited to item title, item model and other information, so as to generate more accurate and richer scenes. In this step, two types of matching relationships are mainly mined: one is the same model matching relationship mining, and the main method is to correlate through the item model attributes to obtain a more common sense and more accurate matching relationship; the second is the applicable item mining, and the main method is to match through the item title to obtain the applicable item corresponding to a certain item.

[0106] In a reference embodiment of the present invention, based on model matching and item main matching, generating a fourth sub-scenario according to item attribute data may include: step S10151 and step S10152.

[0107] Step S10151: According to the item category data, main category items and accessory category items are obtained, and then according to the model attribute data of the main category items, items that match the main category items are selected from the accessory category items to generate a fifth sub-scenario.

[0108] The main purpose of this step is to mine seed scenarios that are in line with common sense and have strong interpretability. There is a strong application demand for matching the same model. This step makes full use of the model attributes provided by the platform for items to mine all possible matching relationships between the same models. This step is limited to items with model attributes, usually 3C category items. The specific implementation can be:

[0109] (i) Select main category items and accessory category items based on the primary category, so as to construct a key-value recommendation vocabulary with the main category items as the key and the accessory category items as the value. The main method is to divide the primary category into the main item category and the accessory item category based on the primary category information provided by the e-commerce platform. The main items may include computer complete machines, mobile phones, notebooks and other items, while the accessory items may include computer-related accessories, mobile phone-related accessories, notebook-related accessories and other items. The relationship between the two is that when users browse and purchase the main items, they can find the accessory items that match the main items through the vocabulary. For example, when a user purchases a certain model of mobile phone, the vocabulary can be used to obtain mobile phone films, protective cases, etc. suitable for this model of mobile phone.

[0110] (II) Classify items with the same model attributes into one category and treat them as separate scenarios. Based on the above division of main category items and accessory category items, extract the model attributes of the items in the main category, and extract the applicable model attributes of the items in the accessory category. Items with the same model attributes of the main items and the applicable model attributes of the accessory items are divided into a set, and at the same time, the main items are used as keys and the accessory items are used as values ​​to integrate them into a key-value vocabulary. Each record in the vocabulary can be considered as a seed scenario.

[0111] Step S10152: Obtain the main core word corresponding to the item library, select the applicable item corresponding to the main core word according to the item title information in the item library, and then establish a corresponding relationship between the main item corresponding to the main core word and the applicable item to generate a sixth seed scenario.

[0112] The main purpose of this step is to mine suitable products for a certain item. For example, for the item red wine, the main purpose of this step is to mine all suitable items for red wine, including red wine stoppers, ice cubes, red wine cabinets and other suitable products. The specific implementation can be:

[0113] (I) Obtain the main core words corresponding to the item library. Based on the mapping table between items and core words, determine the core words of each item in the item library. On this basis, count the core words whose total number of covered items exceeds a certain proportion (can be but not limited to 80%) of the total number of items in the item library, and integrate them into a core word list as the main core words.

[0114] (ii) In the item library, match items whose title information contains the main core word but whose core word is not the main core word. The main purpose of this step is to identify suitable items for which the main core word is applicable. For example, red wine is the main core word, and items whose titles contain red wine but whose core word is not red wine are matched from the item library. Table 3 shows the suitable items found for red wine.

[0115]

[0116] Table 3

[0117] (III) Integrate the main items and applicable items corresponding to the main core words. The main purpose of this step is to generate a key-value word list, where the key is the main item corresponding to the main core word, and the value is the applicable items that match the main core word obtained by the above mining. Each record in the word list can be considered as a seed scene, and an example is shown in Table 4.

[0118]

[0119]

[0120] Table 4

[0121] Figure 4 is a schematic diagram of the main steps of a method for generating a seed scene set based on user behavior data and item attribute data according to an embodiment of the present invention. Figure 4 As shown, the main steps of the method for generating a seed scene set based on user behavior data and item attribute data may include:

[0122] Step S401, obtaining user order data, user search data, user browsing data and item attribute data;

[0123] Step S402, obtaining order items corresponding to the user order data, combining the order items, and generating candidate item matching pairs;

[0124] Step S403, based on the association rules, the confidence and support of the candidate item pair are calculated, and then the item pair corresponding to the user order data is selected from the candidate item pair in combination with the preset association threshold;

[0125] Step S404, upgrading the selected item collocation pair to a core word collocation pair, and then expanding the core word collocation pair using the item library to obtain a final item collocation pair;

[0126] Step S405, clustering and integrating the final item pairings based on a clustering algorithm to generate a first seed scene;

[0127] Step S406, screening the search terms corresponding to the user search data to obtain a set of candidate search terms;

[0128] Step S407, based on the bag-of-words vector, calculate the similarity between the candidate search words in the candidate search word set, and then determine a similar search word set in combination with a preset similarity threshold;

[0129] Step S408, obtaining clicked item data corresponding to a similar search word set;

[0130] Step S409, filtering the clicked item data, and using the filtered clicked item data to generate a second seed scene;

[0131] Step S410, obtaining a user browsing sequence corresponding to the user browsing data;

[0132] Step S411, filtering the user browsing sequence corresponding to the user browsing data based on the sequence length rule and the sequence category rule to obtain a first user browsing sequence;

[0133] Step S412, using the first user browsing sequence to train a click rate prediction model, filtering the first user browsing sequence based on the click rate prediction model to obtain a second user browsing sequence;

[0134] Step S413, determining the vector information corresponding to the browsing sequence of the second user, and then clustering and integrating the browsing sequence of the second user in combination with a clustering algorithm to obtain similar browsing sequences;

[0135] Step S414, generating a third sub-scene according to the items corresponding to the similar browsing sequence;

[0136] Step S415, obtaining main category items and accessory category items according to the item category data, and then selecting items that match the main category items from the accessory category items according to the model attribute data of the main category items, so as to generate a fifth sub-scenario;

[0137] Step S416, obtaining the main core word corresponding to the item library, selecting the applicable item corresponding to the main core word according to the item title information in the item library, and then establishing a corresponding relationship between the main item corresponding to the main core word and the applicable item to generate a sixth sub-scenario;

[0138] Step S417, using the fifth sub-scenario and the sixth sub-scenario to form a fourth sub-scenario;

[0139] Step S418: Use the first seed scene, the second seed scene, the third seed scene and the fourth seed scene to form a seed scene set.

[0140] It should be noted that the execution order of step S415 and step S416 can be adjusted according to the actual situation, and can be executed simultaneously, or step S416 can be executed first and then step S415. In the method for generating a seed scene set in the embodiment of the present invention, the user order data, user search data, user browsing data and item attribute data can be analyzed respectively to generate corresponding seed scenes, and then the generated seed scenes are used to form a seed scene set, which can analyze user behavior data and item attribute data to generate a seed scene set, so as to cover as many item collocation combinations under the seed scene as possible, solving the problem that the prior art cannot fully utilize item content information and user behavior information, and also solving the problem that the number of item collocation combinations is small and the coverage category is insufficient.

[0141] There is a problem of poor interpretability in the item recommendation methods of the prior art. A classic example is that a certain mall once used association rule analysis to mine item combinations based on user common purchase records. An example is that beer and diapers are mutually matched. Such unexplainable combination cases often appear in association rule analysis technology. However, on e-commerce platforms, unexplainable item combination recommendations are very likely to cause user complaints, so this method cannot be directly used on e-commerce platforms. In order to solve this problem, in an embodiment of the present invention, the scene vector of each seed scene in the seed scene set can also be determined, which is equivalent to extracting keywords for each item combination as the name of the item combination, thereby increasing the interpretability of the item combination, which is of great significance for the actual landing of e-commerce platforms.

[0142] Considering that the method of determining the scene vector of the seed scene needs to combine the word segmentation vector library and the item vector library to obtain the specific vector, the construction method of the word segmentation vector library and the item vector library is first explained. Among them, the word segmentation vector library is to segment the item titles in the item library, and then calculate the vector of each segmentation, and then generate the word segmentation vector library; the item vector library is to calculate the vector of each item in the item library, and then generate the item vector library.

[0143] In a reference embodiment of the present invention, the item matching method may further include: performing word segmentation processing on item titles in the item library to obtain item title segmentations; based on an unsupervised word vector algorithm, vectorizing the item title segmentations to obtain the segmentation vectors of the item title segmentations to construct a segmentation vector library; for each item in the item library, obtaining the segmentation vector corresponding to each item, calculating the segmentation vector weight corresponding to each item, and calculating the item vector of each item according to the segmentation vector corresponding to each item and the segmentation vector weight corresponding to each item to construct an item vector library.

[0144] The main purpose of this step is to represent all items in the item library in a vectorized way, and the word segmentation of all items is also represented in a vectorized way. The main method can be: first, the item title of each item in the item library is segmented by a commonly used word segmentation tool, and the word segmentation results of the item title are sorted into a word segmentation sequence, and then the commonly used unsupervised word vector tool is used to vectorize the item title segmentations corresponding to all items in the item library, so that each word segmentation has its own vector, so that a word segmentation vector library can be constructed. Then, for each word segmentation of the item title of each item, its corresponding vector is obtained, and the frequency normalized value of each word segmentation is used as the weight to calculate the vector of the item title, so as to obtain the vector representation of each item, and the detailed examples are shown in Table 5. Among them, word segmentation tools include but are not limited to jieba word segmentation, Harbin Institute of Technology LTP and other tools; unsupervised word vector tools include but are not limited to word2vec, fasttext and other tools.

[0145]

[0146]

[0147] Table 5

[0148] In a reference embodiment of the present invention, determining the scene vector of a seed scene set may include: for each seed scene in the seed scene set, determining the scene vector of each seed scene according to the following method: performing word segmentation processing on the item title under each seed scene to obtain the scene segmentation of each seed scene; calculating the score of the scene segmentation, and then selecting the scene keyword of each seed scene from the scene segmentation according to the score of the scene segmentation according to preset keyword selection conditions; obtaining the keyword vector of the scene keyword from the word segmentation vector library, and calculating the global vector of each seed scene according to the keyword vector; obtaining the item vector of each seed scene and the global vector of each seed scene, and calculating the scene vector of each seed scene.

[0149] The purpose of this step is to calculate the keywords of each seed scene in the seed scene dimension. On the one hand, it can be used for scene explanation, and on the other hand, the keywords are vectorized to obtain the global vector of the seed scene. First, the score of each word segmentation of each item title in each seed scene is calculated, and the N words with the largest scores are taken as the keywords of the seed scene. The specific calculation method is:

[0150] (1) First, calculate the conditional probability of a certain item title segmentation word W appearing under the premise of a certain seed scene S. The formula is as follows:

[0151]

[0152] Among them, n(item title segmentation word W, seed scene S) represents the frequency of title segmentation word W in seed scene S, and n(seed scene S) represents the number of all segmentations in seed scene S.

[0153] (2) Then calculate the conditional probability of a seed scene S appearing under the premise that a certain item title segmentation word W appears. The formula is as follows:

[0154]

[0155] Among them, n (item title segmentation word W) indicates the number of times the segmentation word W appears in the entire item library.

[0156] (3) Finally, the score of each word in each scenario is calculated using the following formula. The larger the score of the word, the more it can represent the keyword of the entire seed scenario:

[0157] score(item title segmentation W) =

[0158] P(item title segmentation W|seed scene S)*P(seed scene S|item title segmentation W)

[0159] In summary, we can get the score of each title word in each seed scene. We can take the top N words from largest to smallest as the keywords of the seed scene. On the one hand, we can give a semantic explanation for the seed scene. On the other hand, we can use the constructed word vector library to map the keywords of the seed scene into keyword vectors. Then, we can weight the keyword vectors to get the global vector of the seed scene. The specific formula is as follows:

[0160]

[0161] The constructed item vector library is used to obtain the item vector contained in the seed scene. Then, the obtained item vector of the seed scene and the calculated global vector of the seed scene are used. The two represent the item information of the seed scene and the global information of the seed scene respectively. The two vector information are weighted averaged to obtain the final scene vector of the seed scene. The specific formula is as follows:

[0162]

[0163] In the above formula, n(item|seed scene S) represents the total number of items in the seed scene S. The entire formula represents the weighted average of all item vectors in the seed scene S, and then superimposed with the global seed scene vector, and finally the weighted average is obtained to obtain the scene vector of the seed scene.

[0164] Figure 5 FIG. 1 is a schematic diagram of the main process of a method for determining a scene vector of a seed scene according to an embodiment of the present invention. Figure 5As shown, the main process of the method for determining a scene vector of a seed scene set may include:

[0165] Step S501, performing word segmentation processing on the item titles in the item library to obtain the item title word segments;

[0166] Step S502: Based on an unsupervised word vector algorithm, vectorize the item title word segmentation to obtain the word segmentation vector of the item title word segmentation to construct a word segmentation vector library;

[0167] Step S503, for each item in the item library, obtain the word segmentation vector corresponding to each item, calculate the word segmentation vector weight corresponding to each item, and calculate the item vector of each item according to the word segmentation vector corresponding to each item and the word segmentation vector weight corresponding to each item, so as to construct an item vector library;

[0168] Step S504, performing word segmentation processing on the item title under the seed scene to obtain scene segmentation of the seed scene;

[0169] Step S505, calculating the scores of the scene segmentation words, and then selecting scene keywords for each seed scene from the scene segmentation words according to the preset keyword selection conditions and the scores of the scene segmentation words;

[0170] Step S506, obtaining a keyword vector of the scene keyword from the word segmentation vector library, and calculating a global vector of the seed scene based on the keyword vector;

[0171] Step S507, obtaining the object vector of the seed scene and the global vector of the seed scene, and calculating the scene vector of the seed scene.

[0172] It should be noted that steps S501 to S503 are used to construct a word segmentation vector library and an item vector library, and can be performed before step S505. After obtaining the scene vector of each seed scene, the scene vector of the seed scene set can be obtained, so that the scene vector of the seed scene set can be used to select the target seed scene corresponding to the target item, and then the target seed scene label to which the target item is marked can be marked. Therefore, in a reference embodiment of the present invention, based on the scene vector of the seed scene set, the target seed scene label to which the target item is marked can include: step S1021, obtaining the item vector of the target item from the item vector library, and calculating the vector distance between the item vector of the target item and the scene vector of each seed scene; step S1022, based on a preset vector distance threshold, according to the vector distance, selecting the target seed scene corresponding to the target item from the seed scene set; step S1023, marking the target seed scene as the target seed scene label to which the target item belongs.

[0173] After obtaining the target object, the constructed object vector library can be used to obtain the object vector corresponding to the target object, and then the scene vector of each seed scene is combined to calculate the vector distance between the target object and each seed scene. The vector distance may include but is not limited to cosine distance and Euclidean distance. After calculating the vector distance between the target object and each seed scene, all seed scenes that meet the preset vector distance threshold can be marked as the seed scene label of the target object by presetting the vector distance threshold, thereby achieving the effect of marking the corresponding seed scene label for any object. In addition, it can be seen that one object can correspond to multiple seed scene labels.

[0174] In the embodiment of the present invention, a seed scene set and a scene vector of the seed scene set can be generated based on user behavior data and item attribute data. However, due to the increasing number of items on the e-commerce platform and the limited number of users on the e-commerce platform, the mined seed scenes cannot cover all items. Therefore, in the embodiment of the present invention, the scene vector of the seed scene and the item vector of the item can be used to mark the item with a seed scene tag, thereby realizing the classification of any item into the existing seed scene, greatly improving the coverage of the seed scene. In addition, the embodiment of the present invention proposes a specific method for obtaining keywords of the seed scene, thereby realizing the interpretability of the seed scene, which is of great significance for recommendation applications on the e-commerce platform.

[0175] After marking the target seed scene label for the target item, that is, after determining the target seed scene corresponding to the target item, the target matching item that matches the target item can be obtained, thereby increasing the user's clicks and improving the user experience. In a reference embodiment of the present invention, according to the target seed scene label, obtaining the target matching item that matches the target item can include: step S1031, obtaining an item belonging to the target seed scene label from the item library, and determining that the obtained item is a first candidate matching item; step S1032, according to a preset attribute filtering rule, selecting a second candidate matching item with the same attribute as the target item from the first candidate matching item; step S1033, based on the trained matching relationship recognition model, selecting a target matching item that has a matching relationship with the target item from the second candidate matching items.

[0176] The main purpose is to generate matching items that meet the characteristics of a given target item. The target item can be a sequence of items that users browse, add to cart, or purchase, or a sequence of items that need matching suggestions when setting up item packages on an e-commerce platform. It should be noted that the items in the item sequence must already exist in the item library of the e-commerce platform. After obtaining the target seed scene corresponding to the target item, all items belonging to the target seed scene can be obtained in the item library of the e-commerce platform, and thus used as the first candidate matching items with matching relationships corresponding to the target item.

[0177] After obtaining the first candidate matching items, the second candidate matching items with the same attributes as the target item can be selected from the first candidate matching items according to the preset attribute filtering rules. First, the attribute information of the item is obtained, where the attribute information includes the applicable age group, applicable gender, applicable model, applicable season, applicable scene, commodity origin, commercial use, style, material, style, function and other attribute information of the item, as shown in Table 6.

[0178]

[0179]

[0180] Table 6

[0181] After obtaining the attribute information of the item, the first candidate matching items under the target seed scene label corresponding to the target item can be filtered according to the preset attribute filtering rules based on the item attributes of the target item. The preset attribute filtering rules may include but are not limited to: the target item and the first candidate matching item must be suitable for the same age group, the target item and the first candidate matching item must be suitable for the same gender, the target item and the first candidate matching item must have the same or similar style, etc.

[0182] In the embodiment of the present invention, a target matching item that has a matching relationship with the target item can also be selected from the second candidate matching items based on the trained matching relationship recognition model. The matching relationship recognition model refers to a model that recognizes whether there is a matching relationship between any two items, and thus can be used to filter the second candidate matching items to obtain the final target matching item.

[0183] As a reference embodiment of the present invention, before selecting a target matching item that has a matching relationship with the target item from the second candidate matching items based on the trained matching relationship recognition model, the item matching method may also include: obtaining user order data and re-behavior data corresponding to the user order data; using a learning algorithm to train the matching relationship recognition model based on the user order data and the re-behavior data. The re-behavior data corresponding to the user order data refers to the behavior data that occurs again after the user order data is submitted. For example, if the user order data is the user's purchase behavior, the re-behavior data may be the user's behavior data after the user purchases the item.

[0184] The specific training method of the collocation relationship recognition model can be: first obtain the re-behavior data corresponding to the user order data, that is, obtain the behavior information after the user order data, such as the user's post-purchase behavior information, which may include behavior information such as add-on purchase and browsing. This information can be obtained from the user behavior information recorded in the e-commerce platform. It is known that the user order data includes A, and after the user submits the order data, he browses items B, C, and D, and purchases items B and E. Then, the data can be organized into a training set. The assumption based on the model is that when the user submits the order data including item A, browses or purchases item B, then item A and item B belong to a collocation relationship. Through large-scale training of the browsing information after the user submits the order data and the purchase information after the user submits the order data, a collocation relationship recognition model can be obtained. The collocation relationship recognition model is a classification model, which can be based on common machine learning algorithms and deep learning algorithms, including but not limited to GBDT models, logistic regression models, deep learning classification models, etc.

[0185] After the matching relationship recognition model is trained, the second candidate matching items can be filtered based on the matching relationship recognition model, and all target items and target matching items predicted to be actual matching relationships are retained, so as to obtain the target matching items corresponding to the final target item and having an accurate matching relationship with it. The final output example is shown in Table 7.

[0186]

[0187]

[0188] Table 7

[0189] In addition, in a reference embodiment of the present invention, after selecting a target matching item that has a matching relationship with the target item from the second candidate matching items, the item matching method may further include: setting an item sorting rule, and sorting the target matching items according to the item sorting rule.

[0190] The main purpose of this step is to sort the acquired target collocation items and put the items with stronger collocation relationship at the front of the sequence. The sorting basis can be provided by the demand side. For example, if the demand side needs to improve the click-through rate of online business, the item sequence can be sorted according to the historical click situation of the items, so as to obtain the collocation product sequence that meets the demand side. The specific implementation method can be: (1) First, obtain the relevant sorting basis based on different businesses. The main purpose of this step is to obtain the sorting basis. On the one hand, under the guarantee of the model and rules, the collocation relationship of the target collocation items of the target item meets the requirements and ensures the user experience. On the other hand, the sequence of the target collocation items is re-sorted in combination with the sorting basis of the demand side, which can ensure that the output item sequence is more in line with the business characteristics of the demand side, thereby improving the business click and conversion indicators. Among them, the sorting basis can include but is not limited to the historical click information of the item, the historical evaluation information of the item, the historical purchase information of the item, etc. (2) Sort and output the final collocation sequence. According to the sorting basis provided in step (1), the acquired target collocation items are sorted to obtain the final sorted sequence of the target collocation items. Taking the sorting of the historical click rate of the item as an example, the target collocation item sequence shown in Table 8 is obtained.

[0191]

[0192]

[0193] Table 8

[0194] Figure 6 FIG. 1 is a schematic diagram of the main process of a method for obtaining a target matching item that matches a target item according to an embodiment of the present invention. Figure 6 As shown, the main process of the method for obtaining a target matching item that matches the target item may include:

[0195] Step S601, obtaining user order data and re-behavior data corresponding to the user order data;

[0196] Step S602, using a learning algorithm to train a matching relationship recognition model based on user order data and repeat behavior data;

[0197] Step S603, obtaining an item belonging to the target seed scene label from the item library, and determining that the obtained item is the first candidate matching item;

[0198] Step S604, selecting second candidate matching items having the same attributes as the target item from the first candidate matching items according to a preset attribute filtering rule;

[0199] Step S605, selecting a target matching item that has a matching relationship with the target item from the second candidate matching items based on the trained matching relationship recognition model;

[0200] Step S606, setting an item sorting rule, and sorting the target matching items according to the item sorting rule.

[0201] In the method for obtaining target matching items that match the target items in the embodiment of the present invention, the matching items that are first obtained are filtered and screened using preset attribute filtering rules and a trained matching relationship recognition model to obtain target matching items, and sorting rules for the target matching items are set according to specific needs, thereby improving user experience and matching accuracy and solving the problem of insufficient accuracy in item matching combinations.

[0202] Figure 7 Schematic diagram of the main modules of the item matching device according to an embodiment of the present invention. Figure 7 As shown, the main modules of the item matching device 700 may include: a generating module 701 , a marking module 702 and an acquiring module 703 .

[0203] Among them, the generation module 701 can be used to: generate a seed scene set based on user behavior data and item attribute data, and determine the scene vector of the seed scene set; the marking module 702 can be used to: obtain the target item, and based on the scene vector of the seed scene set, mark the target item with a target seed scene label; the acquisition module 703 can be used to: obtain the target matching item that matches the target item based on the target seed scene label.

[0204] In the embodiment of the present invention, the generation module 701 can also be used to: obtain user order data, user search data, user browsing data and item attribute data; mine item matching pairs corresponding to the user order data to generate a first seed scene corresponding to the user order data; determine click data corresponding to the user search data to generate a second seed scene corresponding to the user search data; fuse similar browsing sequences corresponding to the user browsing data to generate a third seed scene corresponding to the user browsing data; generate a fourth seed scene based on model matching and item main matching according to item attribute data. Among them, the first seed scene, the second seed scene, the third seed scene and the fourth seed scene constitute a seed scene set.

[0205] In the embodiment of the present invention, the generation module 701 can also be used to: obtain order items corresponding to user order data, combine the order items, and generate candidate item matching pairs; based on the association rules, calculate the confidence and support of the candidate item matching pairs, and then select the item matching pairs corresponding to the user order data from the candidate item matching pairs in combination with a preset association threshold; promote the selected item matching pairs to core word matching pairs, and then use the item library to expand the core word matching pairs to obtain the final item matching pairs; cluster and integrate the final item matching pairs based on the clustering algorithm to generate the first seed scene.

[0206] In an embodiment of the present invention, the generation module 701 can also be used to: filter the search terms corresponding to the user search data to obtain a candidate search term set; calculate the similarity between the candidate search terms in the candidate search term set based on the bag-of-words vector, and then determine the similar search term set in combination with a preset similarity threshold; obtain the clicked item data corresponding to the similar search term set; filter the clicked item data, and use the filtered clicked item data to generate a second seed scene.

[0207] In an embodiment of the present invention, the generation module 701 can also be used to: obtain a user browsing sequence corresponding to the user browsing data; filter the user browsing sequence corresponding to the user browsing data based on a sequence length rule and a sequence category rule to obtain a first user browsing sequence; use the first user browsing sequence to train a click-through rate prediction model, filter the first user browsing sequence based on the click-through rate prediction model to obtain a second user browsing sequence; determine the vector information corresponding to the second user browsing sequence, and then cluster and integrate the second user browsing sequence in combination with a clustering algorithm to obtain a similar browsing sequence; generate a third seed scene based on the items corresponding to the similar browsing sequences.

[0208] In the embodiment of the present invention, the generation module 701 can also be used to: obtain main category items and accessory category items according to the item category data, and then select items that match the main category items from the accessory category items according to the model attribute data of the main category items to generate the fifth seed scene; obtain the main core word corresponding to the item library, select the applicable items corresponding to the main core word according to the item title information in the item library, and then establish the corresponding relationship between the main item corresponding to the main core word and the applicable item to generate the sixth seed scene. Among them, the fifth seed scene and the sixth seed scene constitute the fourth seed scene.

[0209] In an embodiment of the present invention, the generation module 701 can also be used to: for each seed scene in the seed scene set, determine the scene vector of each seed scene according to the following method: perform word segmentation processing on the item title under each seed scene to obtain the scene segmentation of each seed scene; calculate the score of the scene segmentation, and then select the scene keyword of each seed scene from the scene segmentation according to the preset keyword selection condition and the score of the scene segmentation; obtain the keyword vector of the scene keyword from the word segmentation vector library, and calculate the global vector of each seed scene according to the keyword vector; obtain the item vector of each seed scene and the global vector of each seed scene, and calculate the scene vector of each seed scene.

[0210] In an embodiment of the present invention, the marking module 702 can also be used to: obtain the item vector of the target item from the item vector library, and calculate the vector distance between the item vector of the target item and the scene vector of each seed scene; based on a preset vector distance threshold, select the target seed scene corresponding to the target item from the seed scene set according to the vector distance; and mark the target seed scene as the target seed scene label to which the target item belongs.

[0211] In the embodiment of the present invention, the item matching device may further include a construction module (not shown in the figure). The construction module may be used to: perform word segmentation processing on the item titles in the item library to obtain the item title word segmentation; based on the unsupervised word vector algorithm, perform vectorization processing on the item title word segmentation to obtain the word segmentation vector of the item title word segmentation to construct a word segmentation vector library; for each item in the item library, obtain the word segmentation vector corresponding to each item, calculate the word segmentation vector weight corresponding to each item, and calculate the item vector of each item according to the word segmentation vector corresponding to each item and the word segmentation vector weight corresponding to each item to construct an item vector library.

[0212] In the embodiment of the present invention, the acquisition module 703 can also be used to: acquire items belonging to the target seed scene label from the item library, and determine that the acquired items are first candidate matching items; select second candidate matching items with the same attributes as the target items from the first candidate matching items according to preset attribute filtering rules; and select target matching items that have a matching relationship with the target item from the second candidate matching items based on the trained matching relationship recognition model.

[0213] In an embodiment of the present invention, the acquisition module 703 can also be used to: obtain user order data and repeated behavior data corresponding to the user order data; use a learning algorithm to train a matching relationship recognition model based on the user order data and the repeated behavior data; and set item sorting rules to sort the target matching items according to the item sorting rules.

[0214] From the above description, it can be seen that the item matching device of the embodiment of the present invention can analyze the user behavior data and item attribute data to generate a seed scene set, so as to cover as many item matching combinations as possible under the seed scene, solving the problem that the prior art cannot fully utilize the item content information and user behavior information, and also solves the problem of a small number of item matching combinations and insufficient coverage of categories, thereby improving user experience and matching accuracy. In addition, in the embodiment of the present invention, a scene vector can also be determined, which is equivalent to extracting keywords for each item matching combination as the name of the item matching combination, thereby increasing the interpretability of the item matching and solving the problem of the interpretability of the item matching combination, which is of great significance for the actual implementation of the e-commerce platform.

[0215] Figure 8An exemplary system architecture 800 is shown to which the item collocation method or item collocation device according to the embodiment of the present invention can be applied.

[0216] like Figure 8 As shown, system architecture 800 may include terminal devices 801, 802, 803, a network 804, and a server 805. Network 804 is used to provide a medium for communication links between terminal devices 801, 802, 803 and server 805. Network 804 may include various connection types, such as wired, wireless communication links, or optical fiber cables, etc.

[0217] Users can use terminal devices 801, 802, and 803 to interact with server 805 through network 804 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 801, 802, and 803, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (only examples).

[0218] The terminal devices 801 , 802 , and 803 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0219] The server 805 may be a server that provides various services, such as a backend management server (only an example) that provides support for shopping websites browsed by users using the terminal devices 801, 802, and 803. The backend management server may analyze and process the received data such as product information query requests, and feed back the processing results (such as target push information, product information - only an example) to the terminal device.

[0220] It should be noted that the item matching method provided in the embodiment of the present invention is generally executed by the server 805 , and accordingly, the item matching device is generally disposed in the server 805 .

[0221] It should be understood that Figure 8 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0222] Reference below Fig. 9 , which shows a schematic diagram of the structure of a computer system 900 of a terminal device suitable for implementing an embodiment of the present invention. Fig. 9 The terminal device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0223] like Fig. 9As shown, the computer system 900 includes a central processing unit (CPU) 901, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage part 908 into a random access memory (RAM) 903. In the RAM 903, various programs and data required for the operation of the system 900 are also stored. The CPU 901, the ROM 902, and the RAM 903 are connected to each other via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0224] The following components are connected to the I / O interface 905: an input section 906 including a keyboard, a mouse, etc.; an output section 907 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card, a modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to the I / O interface 905 as needed. A removable medium 911, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 910 as needed, so that a computer program read therefrom is installed into the storage section 908 as needed.

[0225] In particular, according to the embodiments disclosed in the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 909, and / or installed from the removable medium 911. When the computer program is executed by the central processing unit (CPU) 901, the above-mentioned functions defined in the system of the present invention are executed.

[0226] It should be noted that the computer-readable medium shown in the present invention may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present invention, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0227] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, functions and operations of the system, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0228] The modules involved in the embodiments of the present invention may be implemented in software or hardware. The modules described may also be set in a processor. For example, they may be described as: a processor includes a generation module, a marking module, and an acquisition module. The names of these modules do not, in some cases, constitute limitations on the modules themselves. For example, the generation module may also be described as a "module for generating a seed scene set based on user behavior data and item attribute data, and determining a scene vector of the seed scene set."

[0229] As another aspect, the present invention further provides a computer-readable medium, which may be included in the device described in the above embodiment; or may exist independently without being assembled into the device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by a device, the device includes: generating a seed scene set based on user behavior data and item attribute data, determining a scene vector of the seed scene set; obtaining a target item, and based on the scene vector of the seed scene set, marking a target seed scene tag for the target item; and obtaining a target matching item that matches the target item based on the target seed scene tag.

[0230] According to the technical solution of the embodiment of the present invention, it is possible to analyze user behavior data and item attribute data to generate a seed scene set, so as to cover as many item combinations as possible under the seed scene, solving the problem that the prior art cannot fully utilize item content information and user behavior information, and also solves the problem of a small number of item combinations and insufficient coverage of categories, thereby improving user experience and matching accuracy. In addition, in the embodiment of the present invention, a scene vector can also be determined, which is equivalent to extracting keywords for each item combination as the name of the item combination, thereby increasing the interpretability of the item combination and solving the problem of the interpretability of the item combination, which is of great significance for the actual implementation of the e-commerce platform.

[0231] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for matching items, characterized in that: include: Generate a seed scene set according to the user behavior data and the item attribute data, and determine the scene vector of the seed scene set, including: Obtain user order data, user search data, user browsing data and item attribute data; Mining item pairings corresponding to the user order data to generate a first seed scene corresponding to the user order data; Determine click data corresponding to the user search data, and generate a second seed scenario corresponding to the user search data; The similar browsing sequences corresponding to the user browsing data are integrated to generate a third seed scenario corresponding to the user browsing data, including: obtaining a user browsing sequence corresponding to the user browsing data; filtering the user browsing sequence corresponding to the user browsing data based on a sequence length rule and a sequence category rule to obtain a first user browsing sequence; using the first user browsing sequence to train a click rate prediction model, filtering the first user browsing sequence based on the click rate prediction model to obtain a second user browsing sequence; determining vector information corresponding to the second user browsing sequence, and then clustering and integrating the second user browsing sequence in combination with a clustering algorithm to obtain the similar browsing sequence; generating the third seed scenario according to the items corresponding to the similar browsing sequence; Based on the model matching and the item main matching, a fourth seed scene is generated according to the item attribute data; wherein the first seed scene, the second seed scene, the third seed scene and the fourth seed scene constitute the seed scene set; Acquire a target object, and based on the scene vector of the seed scene set, label the target object with a target seed scene tag; According to the target seed scene label, a target matching item that matches the target item is obtained.

2. The method according to claim 1, characterized in that The mining of item pairings corresponding to the user order data to generate a first seed scenario corresponding to the user order data includes: Obtaining order items corresponding to the user order data, combining the order items, and generating candidate item matching pairs; Based on the association rules, the confidence and support of the candidate item pair are calculated, and then the item pair corresponding to the user order data is selected from the candidate item pair in combination with a preset association threshold; The selected item collocation pair is promoted to a core word collocation pair, and then the core word collocation pair is expanded by using an item library to obtain a final item collocation pair; The final item pairings are clustered and integrated based on a clustering algorithm to generate the first seed scene.

3. The method according to claim 1, characterized in that The determining the click data corresponding to the user search data and generating the second seed scenario corresponding to the user search data includes: Screening the search terms corresponding to the user search data to obtain a set of candidate search terms; Based on the bag-of-words vector, the similarity between the candidate search words in the candidate search word set is calculated, and then a similar search word set is determined in combination with a preset similarity threshold; Obtaining clicked item data corresponding to the similar search term set; The clicked object data is filtered, and the second seed scene is generated using the filtered clicked object data.

4. The method according to claim 1, characterized in that: The generating of the fourth sub-scenario based on the model matching and the item main matching according to the item attribute data includes: According to the item category data, main category items and accessory category items are acquired, and then according to the model attribute data of the main category items, items that match the main category items are selected from the accessory category items to generate a fifth sub-scenario; Obtain the main core word corresponding to the item library, select the applicable item corresponding to the main core word according to the item title information in the item library, and then establish a corresponding relationship between the main item corresponding to the main core word and the applicable item to generate a sixth sub-scenario; wherein, The fifth sub-scenario and the sixth sub-scenario constitute the fourth sub-scenario.

5. The method according to claim 1, characterized in that The determining of the scene vector of the seed scene set includes: For each seed scene in the seed scene set, the scene vector of each seed scene is determined according to the following method: Performing word segmentation processing on the item title under each seed scene to obtain scene segmentation of each seed scene; Calculating the scores of the scene segmentations, and then selecting scene keywords for each seed scene from the scene segmentations according to the scores of the scene segmentations according to preset keyword selection conditions; Obtaining a keyword vector of the scene keyword from a word segmentation vector library, and calculating a global vector of each seed scene according to the keyword vector; Obtain the object vector of each seed scene and the global vector of each seed scene, and calculate the scene vector of each seed scene.

6. The method according to claim 5, characterized in that The scene vector based on the seed scene set is a target seed scene label to which the target object tag belongs, including: Obtaining the item vector of the target item from the item vector library, and calculating the vector distance between the item vector of the target item and the scene vector of each seed scene; Based on a preset vector distance threshold, and according to the vector distance, selecting a target seed scene corresponding to the target object from the seed scene set; The target seed scene is marked as the target seed scene tag to which the target object belongs.

7. The method according to claim 6, characterized in that The method further comprises: Perform word segmentation on the item titles in the item library to obtain the item title word segments; Based on an unsupervised word vector algorithm, vectorize the word segmentation of the item title to obtain the word segmentation vector of the item title segmentation to construct the word segmentation vector library; For each item in the item library, obtain the word segmentation vector corresponding to each item, calculate the word segmentation vector weight corresponding to each item, and calculate the item vector of each item according to the word segmentation vector corresponding to each item and the word segmentation vector weight corresponding to each item to construct the item vector library.

8. The method according to claim 1, characterized in that The step of acquiring a target matching item that matches the target item according to the target seed scene label includes: Acquire an item belonging to the target seed scene label from an item library, and determine that the acquired item is a first candidate matching item; According to a preset attribute filtering rule, selecting a second candidate matching item having the same attribute as the target item from the first candidate matching items; Based on the trained collocation relationship recognition model, the target collocation item having a collocation relationship with the target item is selected from the second candidate collocation items.

9. The method according to claim 8, characterized in that Before selecting the target matching item having a matching relationship with the target item from the second candidate matching items based on the trained matching relationship recognition model, the method further includes: Obtaining user order data and re-behavior data corresponding to the user order data; Using a learning algorithm, training a collocation relationship recognition model based on the user order data and the repeat behavior data; and After selecting the target matching item that has a matching relationship with the target item from the second candidate matching items, the method further includes: An item sorting rule is set, and the target matching items are sorted according to the item sorting rule.

10. An item matching device, characterized in that: include: A generation module is used to generate a seed scene set according to user behavior data and item attribute data, and determine a scene vector of the seed scene set, including: Obtain user order data, user search data, user browsing data and item attribute data; Mining item pairings corresponding to the user order data to generate a first seed scene corresponding to the user order data; Determine click data corresponding to the user search data, and generate a second seed scenario corresponding to the user search data; The similar browsing sequences corresponding to the user browsing data are integrated to generate a third seed scenario corresponding to the user browsing data, including: obtaining a user browsing sequence corresponding to the user browsing data; filtering the user browsing sequence corresponding to the user browsing data based on a sequence length rule and a sequence category rule to obtain a first user browsing sequence; using the first user browsing sequence to train a click rate prediction model, filtering the first user browsing sequence based on the click rate prediction model to obtain a second user browsing sequence; determining vector information corresponding to the second user browsing sequence, and then clustering and integrating the second user browsing sequence in combination with a clustering algorithm to obtain the similar browsing sequence; generating the third seed scenario according to the items corresponding to the similar browsing sequence; Based on the model matching and the item main matching, a fourth seed scene is generated according to the item attribute data; wherein the first seed scene, the second seed scene, the third seed scene and the fourth seed scene constitute the seed scene set; A marking module, used to obtain a target object and mark a target seed scene label for the target object based on the scene vector of the seed scene set; An acquisition module is used to acquire a target matching item that matches the target item according to the target seed scene label.

11. An electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 9.

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

Images