Commodity recommendation method and device

Abstract

The invention provides a commodity recommendation method and device, and the method comprises the steps: obtaining a recalled commodity set for each user through employing a multi-path recall mode, and enabling the multi-path recall mode to comprise at least two of label recommendation recall, collaborative filtering recommendation recall, hot recommendation recall, user historical recall and vector recall; a user browsing feature, a user click feature, a user purchase feature, a user customer level feature, a user gender feature and a commodity differentiation feature are constructed, a LightGBM algorithm model is adopted to perform fine ranking on the click rate to obtain a fine ranking sequence for each user, and the fine ranking sequence for each user is scattered according to a second-level category. And forming a mixed arrangement sequence for each user. According to the scheme, the number and types of commodities recommended to the user by the online shopping website are enriched, and the click rate and conversion rate of the recommended commodities are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a commodity recommendation method and device. Background Art

[0002] Home page recommendation is the core scenario for users to browse supermarkets. Usually, the user arrival rate of the home page recommendation scenario is second only to the search page and category page, and higher than other pages. In the prior art, the page view (Page View, PV for short) click-through rate of the store home page recommendation recommended by big data (collaborative filtering) is higher than that of the manually arranged store home page recommendation. For example, the page view PV click-through rate of the manually arranged store home page recommendation is 0.53%, and the click-through conversion rate is 14.32%; while the PV click-through rate of the store home page recommendation recommended by big data (collaborative filtering) is 0.71%, and the click-through conversion rate is 21.55%. The home page big data recommendation has achieved initial results, but it is still far from the PV click-through rate and click-through conversion rate of the search results page. For example, the PV click-through rate of the search results page is 2.5%, and the click-through conversion rate is 40%. Therefore, the construction of the home page recommendation link and the click-through rate and conversion rate obtained still have a lot of room for improvement.

[0003] In order to improve the user's homepage browsing experience and increase the click-through rate of users recommended on the homepage, a page recommendation method is provided, which has great commercial value. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a product recommendation method and device, which can enhance the user's online browsing experience and increase the user click rate of the recommended products.

[0005] In order to solve the above problems, the present invention provides a product recommendation method, including: using a multi-way recall method to obtain a recalled product set for each user, the multi-way recall includes at least two of tag recommendation recall, collaborative filtering recommendation recall, popular recommendation recall, user history recall and vector recall; feature construction, the constructed features include user browsing features, user click features, user purchase features, user customer layer features, user gender features and differentiated features of products; according to the features, using the LightGBM algorithm model to train and predict the click-through rate of the products in the recalled product set of each user, and sorting the products in the recalled product set for each user from large to small according to the predicted click-through rate to obtain a refined sequence for each user; breaking up the refined sequence for each user according to the secondary categories to form a mixed sequence for each user, and using the mixed sequence for each user as the recommendation sequence for each user.

[0006] In some embodiments, the tags selected for the tag recommendation recall include the final category value, customer segment value, brand value, and whether it is a promotion.

[0007] In some embodiments, the collaborative filtering recommendation recall calculates the similarity between products through the user's ordering behavior.

[0008]

[0009] Among them, A and B represent two commodities respectively.

[0010] In some embodiments, the popular recommendation recall sorts the popular products in the community according to click-through rate, conversion rate, number of repeat purchases in the community, and online sales, selects multiple products and puts them into the recalled product set, and uses Bayesian smoothing correction for the click-through rate and conversion rate.

[0011] In some embodiments, the user history recall selects multiple commodities that the user has searched, browsed, and purchased in the history and puts them into the recalled commodity set.

[0012] In some embodiments, the vector recall obtains user vectors and product vectors through a word vector generation model, constructs a deep semantic matching model, trains the deep semantic matching model and predicts the similarity between the user side and the product side, and finally sorts the predicted similarity scores from large to small, selects multiple products and puts them into the recalled product set. When the deep semantic matching model is trained, the label is whether it is clicked.

[0013] In some embodiments, in the step of using the LightGBM algorithm model to train and predict the click-through rate of the products in the recalled product set of each user, the constructed features are screened according to the feature importance of the LightGBM algorithm model, and the LightGBM algorithm model is run in a dynamic partitioned distributed manner.

[0014] In some embodiments, the product click-through rate CTR = N / M, where N is the number of clicks and M is the amount of exposure. The Bayesian smoothing correction formula is:

[0015]

[0016] in, is the corrected CTR, is the mean, S 2 is the variance, Conversion rate = product sales / (number of clicks to view product details + number of clicks to add to cart).

[0017] In some embodiments, the step of breaking up the finely sorted sequence for each user according to the secondary categories to form a mixed sequence for each user further includes: setting a sorted sequence and a sequence to be sorted, the sorted sequence is used to store the products that have been sorted, and the sequence to be sorted is used to store the products that have been taken out of the finely sorted sequence but have not been sorted; looping through the following steps until all the products in the finely sorted sequence are stored in the sorted sequence: selecting the current product, taking a product from the head of the sequence to be sorted as the current product, if the sequence to be sorted is empty or all the products in the sequence to be sorted have the same secondary category as the two products at the end of the sequence in the sorted sequence, then taking a product from the head of the remaining product sequence in the finely sorted sequence as the current product;

[0018] Determine whether the current product and the two products at the end of the sorted column have the same secondary category. If they do, store the current product at the end of the column to be sorted. If they do not, store the current product at the end of the sorted column. After all products are sorted, use the sorted column as the mixed sorting sequence.

[0019] In order to solve the above problems, the present invention also provides a product recommendation device, including: a product recall module: used to obtain a recalled product set for each user by adopting a multi-way recall method, and the product recall module includes at least two of the multi-way recall including a label recommendation recall unit, a collaborative filtering recommendation recall unit, a hot recommendation recall unit, a user history recall unit and a vector recall unit; the label recommendation recall unit is used to perform label recommendation recall on products, the collaborative filtering recommendation recall unit is used to perform collaborative filtering recommendation recall on products, and the user history recall unit is used to perform user history recall on products. , the vector recall unit is used to perform vector recall on commodities; feature construction module: used for feature construction, the constructed features include: user browsing features, click features, purchase features, customer layer features, gender features and differentiated features of commodities; sorting module: including a refined sorting unit and a mixed sorting unit; the refined sorting unit uses the LightGBM algorithm model to train and predict the click-through rate of the commodities in the recalled commodity set of each user according to the constructed features, and sorts the commodities in the recalled commodity set for each user from large to small according to the predicted click-through rate to obtain a refined sorting sequence for each user;

[0020] The shuffling unit breaks up the refined sequence for each user according to secondary categories to form the shuffled sequence for each user, and uses the shuffled sequence for each user as the recommended sequence for each user.

[0021] The above technical solution obtains the recalled product set for each user by adopting a multi-channel recall method, wherein the multi-channel recall includes at least two of tag recommendation recall, collaborative filtering recommendation recall, popular recommendation recall, user history recall and vector recall, constructs user browsing features, user click features, user purchase features, user customer layer features, user gender features and differentiated features of products, and uses the LightGBM algorithm model to finely rank the click-through rate to obtain a finely ranked sequence for each user, and breaks up the finely ranked sequence for each user according to the secondary category to form a mixed sequence for each user. It enriches the number and types of products recommended to users by online shopping websites, and improves the click-through rate and conversion rate of recommended products.

[0022] It should be understood that the above general description and the detailed description below are exemplary and explanatory only and cannot limit the present invention. The techniques, methods and devices known to ordinary technicians in the relevant field may not be discussed in detail, but where appropriate, the techniques, methods and devices should be considered as part of the authorization specification. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments of the present invention are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0024] Figure 1 This is an application scenario diagram of a product recommendation method provided by an embodiment of the present invention;

[0025] Figure 2 is a flow chart of a product recommendation method provided by an embodiment of the present invention;

[0026] Figure 3 It is a schematic diagram of a sequence state formed by the main steps of the shuffling method provided by an embodiment of the present invention;

[0027] Figure 4 It is a schematic diagram of a commodity recommendation device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] Figure 1FIG. 1 is an application scenario diagram of a commodity recommendation method provided by an embodiment of the present invention. Figure 1 As shown, the application scenario includes a server 11, a terminal 12 and a network 13, and the terminal 12 and the server 11 exchange information through the network 13. The terminal 12 can be any terminal such as a desktop computer, a laptop computer, a tablet computer or a mobile phone that can install applications or access web pages through a browser. The network 13 can be a wired network or a wireless network. The server 11 is used to provide background services for the application of the terminal 12. The server 11 can be a single server, or more often a server cluster. In general, the product recommendation method of the present invention is completed on the server 11, and the generated results are pushed to the terminal 12 for display.

[0030] The server 11 includes hardware and programs for running the product recommendation method provided by the present invention. The server 11 also includes a database, which stores necessary data during the execution of the product recommendation method provided by the present invention and a recommendation sequence for each user generated by the product recommendation method provided by the present invention.

[0031] It should be noted that Figure 1 The numbers of terminals, networks, and servers are schematic and are illustrated based on the application scenario recommended on the homepage of a shopping website.

[0032] Combine the following Figure 2-Figure 3 , the commodity recommendation method described in the present invention is described in detail.

[0033] Figure 2 FIG. 1 is a flow chart of a commodity recommendation method provided by an embodiment of the present invention. Figure 2As shown, the product recommendation method includes: step S21, using a multi-way recall method to obtain a recalled product set for each user, the multi-way recall includes at least two of tag recommendation recall, collaborative filtering recommendation recall, popular recommendation recall, user history recall and vector recall; step S22, feature construction, the constructed features include user browsing features, user click features, user purchase features, user customer layer features, user gender features and differentiated features of products; step S23, according to the constructed features, using a lightweight gradient boosting machine (Light Gradient Boosting Machine, referred to as LightGBM) algorithm model to train and predict the click-through rate of the products in the recalled product set of each user, and sorting the products in the recalled product set for each user from large to small according to the predicted click-through rate to obtain a refined sequence for each user; step S24, breaking up the refined sequence for each user according to the secondary category to form a mixed sequence for each user, and using the mixed sequence for each user as the recommendation sequence for each user.

[0034] Referring to step S21, a multi-channel recall method is used to obtain a recalled product set for each user, wherein the multi-channel recall includes at least two of tag recommendation recall, collaborative filtering recommendation recall, popular recommendation recall, user history recall, and vector recall. This step obtains the multi-channel recall results for each user and removes duplicates from the multi-channel recall results to obtain a recalled product set for each user.

[0035] In this step, the labels selected by the label recommendation recall include the final category value, customer value, brand value, and whether it is a promotion. Among them, the final category value refers to the last category of the product on the shopping website, and the customer value includes different customer levels such as high, medium, and low. The label recommendation recall calculates the user's interest value for the label value corresponding to a certain label, and then weights each label value to the product, converts the user's preference for each label into a preference for the product, and calculates the users who have added purchases, browsed, and searched in the past 90 days in the dimension of the community (for example: taking a city as a community), and sorts them by preference score. Each user takes the first 200 items. The specific calculation method is: user's rating of the product = user's interest matrix for the label * transposed matrix of the correlation matrix between the product and the label.

[0036] The elements in the user's interest matrix for tags are composed of the user's interest value for a certain tag value. The user's interest value for a certain tag value = 0.5*behavior type weight in the past 30 days*number of behaviors*TF in the past 30 days*IDF in the past 30 days+0.3*behavior type weight in the past 90 days*number of behaviors*TF in the past 90 days*IDF in the past 90 days+0.2*behavior type weight in the past year*number of behaviors*TF in the past year*IDF in the past year

[0037] Among them, TF represents the user's dependence on a certain tag. The calculation method of TF is: TF = the sum of the number of times the user's behavior hits the tag / the sum of the number of times the user's behavior hits all tags;

[0038] IDF indicates the rarity of the tag. The IDF calculation method is:

[0039] IDF=log 10 (the sum of all users' tag counts for all tags and / the sum of all users' tag counts for that tag + 1).

[0040] In this step, the behavior types include ordering behavior, adding to cart behavior, browsing behavior, and searching behavior, and the corresponding weight values ​​are set to 80, 50, 12.5, and 1, respectively. Among them, ordering behavior, adding to cart behavior, and browsing behavior are used to calculate the user's interest in the final category value, customer value, brand value, whether it is a promotion, and other labels, while searching behavior is only used to calculate the user's interest in the final category value label.

[0041] In this step, for ordering behavior, if the user places an order for the same product multiple times on the same day, the number of ordering behaviors is 1; for adding to cart behavior, if the user adds to cart for the same product multiple times on the same day, the number of adding to cart behavior is 1; the search behavior takes the last-level classification value corresponding to the first product in the search results, and each search for the same product by the user on the same day is counted as 1 search behavior.

[0042] The product and label correlation matrix consists of the correlation matrix between products and different label values.

[0043] In this step, the labels selected for the label recommendation recall include the final category value, customer value, brand value, and whether it is a promotion. A product and label correlation matrix is ​​set according to different labels. For example, when the correlation matrix of the product to the final category value label is set, if the product is in the current customer group, it is recorded as 1, and if it is not in the current customer group, it is recorded as 0; when the correlation matrix of the product to the brand label is set, if it is the current brand, it is recorded as 1, and if it is not the current brand, it is recorded as 0; when the correlation matrix of the product to the promotion label is set, if it is on promotion, it is recorded as 1, and if it is not on promotion, it is recorded as 0; when the correlation matrix of the product to the promotion label is set, if the final category of the product is a second-level category, then the correlation between the product and the final category is recorded as 2, the correlation with the sibling final category (the same first-level category) is recorded as 1, and the correlation with the remaining categories is recorded as 0; if the final category of the product is a third-level category, then the correlation between the product and the final category is recorded as 2, the correlation with the sibling third-level category (the same second-level category) is recorded as 1.5, the correlation with the sibling second-level category and its subordinate final category (the same first-level category) is recorded as 1, and the correlation with the remaining categories is recorded as 0.

[0044] The collaborative filtering recommendation recall recommends similar products to users based on the similarity between products liked by users. The similarity between users and products = the actual liking between users and products * the similarity score of similar products of the product. Users and products are sorted by similarity scores, and each user can take at most the first 200 items.

[0045] First, the similarity of similar products of a product is calculated. Since the user's click behavior is random, the ordering behavior is introduced to enhance the interpretability of the recommendation.

[0046] In this step,

[0047]

[0048] Among them, A and B represent two commodities respectively.

[0049] For example, statistics are collected on the online ordering behaviors of users in the region in the past 90 days (online sales only count orders whose source is online, such as Taoxianda, Ele.me online, ordinary Youxian, Maochao Half-Day Delivery, Domain Half-Day Delivery, and Youxian Half-Day Delivery), and products with less than 20 co-purchased users are filtered out (the similarity is judged to be 0).

[0050] The actual preference between users and products is calculated based on user behaviors. In this step, the selected user behaviors include ordering behavior, adding to cart behavior, and browsing behavior, and the corresponding weight values ​​are set to 80, 50, and 12.5, respectively.

[0051] For users, select the products that have been clicked or ordered in the community in the past 30 days, from recent to oldest. After removing duplicates, select the top 200 products.

[0052] According to the total set of the user's ordering, adding to cart, and browsing behaviors for these 200 products in the past 30 days, t is marked as 1, 2, 3...300 according to the time of the behaviors from recent to distant. For example, there are 3 behaviors in total, the most recent behavior t=1, the second behavior t=2, and the earliest behavior t=3.

[0053] The actual likeness between the user and the product = sum(the weight of the behavior*(e^(-a*t mark of the behavior))*the number of times the behavior occurs), where a is 0.01.

[0054] In this step, the popular recommendation recall sorts the popular products in the community according to the click-through rate, conversion rate, number of repeat purchases in the community, and online sales, selects multiple products and puts them into the recalled product set, and uses Bayesian smoothing correction for the click-through rate and conversion rate. In this embodiment, the click-through rate of the products in the community in the past 30 days is calculated, and the 60 products with the highest click-through rate are selected; the conversion rate of the products in the community in the past 30 days is calculated, and the 60 products with the highest conversion rate are selected; the number of repeat purchases of the products in the community in the past 90 days is calculated, and the 60 products with the highest number of repeat purchases are selected; the online sales of the products in the community in the past 90 days are calculated, and the 30 products with the highest online sales are selected.

[0055] Among them, click rate = CTR = N / M, N is the number of clicks, and M is the number of exposures. In this step, Bayesian smoothing is used to correct the click rate in the popular recall, and the Bayesian smoothing correction formula is:

[0056]

[0057] in, is the corrected CTR, is the mean, S 2 is the variance,

[0058] Conversion rate = product sales / (number of clicks to enter product details + number of clicks to add to cart). The conversion rate can also be corrected using the same method, which will not be repeated here.

[0059] Using Bayesian smoothing to correct the click-through rate and conversion rate in popular recalls can avoid the situation where products with fewer clicks but higher click-through rates are ranked at the top, and avoid the situation where products with fewer purchases but higher conversion rates are ranked at the top.

[0060] Number of repeat purchasers = 0.3*number of people who have ordered a certain product >= 2 times in the past 90 days + 0.7*number of people who have ordered a certain product >= 2 times in the past 30 days. The calculated number of people is deduplicated according to the unique identifier of the user to obtain the number of repeat purchasers. In this step, buying multiple items of the same product in the same order is counted as one time.

[0061] Online sales = 0.3*the online sales of the store in the past 90 days + 0.7*the online sales of the store in the past 30 days. In this embodiment, online sales only count orders from Taoxianda Online, Ele.me Online, Ordinary Youxian, Maochao Half-Day Delivery, Domain Half-Day Delivery, and Youxian Half-Day Delivery.

[0062] In this step, the user history recall selects multiple commodities that the user has searched, browsed, and purchased in the history and puts them into the recalled commodity set.

[0063] In this embodiment, the top three products in the comprehensive ranking of the user's most recent five search terms in the past 90 days are selected to ensure that there are no more than three products in the same final category, for a total of 15 products; the five products with the highest browsing scores of the user in the past 90 days are selected, where browsing score = number of browsing times from 60 to 90 days * 0.2 + number of browsing times from 30 to 60 days * 0.3 + number of browsing times within 30 days * 0.5; the top five products in descending order of the number of orders placed by the user in the past 90 days are selected.

[0064] During the user history recall process, each user takes a maximum of 25 products, which accounts for a relatively small proportion in the recalled product set, and can further increase the richness of new products.

[0065] In this step, the vector recall obtains the user vector and product vector through the word2vec word vector generation model, builds a deep semantic matching model (DSSM), trains the deep semantic matching model and predicts the similarity between the user side and the product side, and finally sorts the predicted similarity scores from large to small, selects multiple products and puts them into the recalled product set. The label of the deep semantic matching model training is whether it is clicked. In the process of vector recall, each user can take up to 200 products.

[0066] The community is an area divided according to market size. In this step, the community refers to a city.

[0067] The use of a multi-channel recall method greatly increases the number of products in the recalled product set for each user, thereby greatly enriching the recommended products for the user. In this embodiment, the multi-channel recall method is used to increase the number of recommended products on the homepage of the shopping website from 100 to 300.

[0068] Referring to step S22, feature construction, the constructed features include user browsing features, user click features, user purchase features, user demographic features, user gender features, and differentiated features of commodities. Among them, user browsing features, user click features, and user purchase features can represent user interests, user demographic features can represent user demographic attributes, and user gender features can represent user gender attributes. Differentiated features of commodities select commodity attribute features, statistical features of commodity browsing, clicking, and purchase, and statistical features of commodity browsing, clicking, and purchase in its sub-category. Differentiated features of commodities can represent commodity features and different characteristics of commodities.

[0069] Referring to step S23, according to the constructed features, the LightGBM algorithm model is used to train and predict the click-through rate of the products in the recalled product set of each user, and the products in the recalled product set for each user are sorted from large to small according to the predicted click-through rate to obtain a refined sort sequence for each user.

[0070] Since the amount of data on users and products is large, the total data flow of the recalled product set for each user is very large, for example, it reaches 4.6 billion in this embodiment. If the LightGBM algorithm model is directly used, the prediction time will be very long.

[0071] In this step, in the step of using the LightGBM algorithm model to train and predict the click rate, the constructed features are screened by the feature importance of the LightGBM algorithm model, and the LightGBM algorithm model is run in a distributed manner using dynamic partitioning. For example, first, the constructed features are screened according to the model feature importance, the number of features is reduced from 246 to 160, dynamic partitioning is used, prediction is performed in groups of 8 million data, and ten nodes are used for distributed parallel computing.

[0072] Referring to step S24, the refined sorting sequence for each user is scattered according to the secondary categories to form a mixed sorting sequence for each user, and the mixed sorting sequence for each user is used as the recommendation sequence for each user. This step can prevent users from continuously seeing the same type of products.

[0073] This step further includes: setting a sorted column and a to-be-sorted column, the sorted column is used to store the products that have been sorted, and the to-be-sorted column is used to store the products that have been taken out of the finely sorted column but have not been sorted; looping the following steps until all the products in the finely sorted column are stored in the sorted column: selecting the current product, taking a product from the head of the column of the to-be-sorted column as the current product, if the to-be-sorted column is empty or all the products in the to-be-sorted column have the same secondary category as the two products at the end of the sorted column, then taking a product from the head of the remaining product sequence in the finely sorted column as the current product; judging whether the current product and the two products at the end of the sorted column have the same secondary category, if so, storing the current product at the end of the column of the to-be-sorted column, if not, storing the current product at the end of the column of the sorted column; after all products are sorted, use the sorted column as the mixed sorting sequence.

[0074] Figure 3 It is a schematic diagram of a sequence formed by the main steps of the shuffling method provided in one embodiment of the present invention. Figure 3In the figure, the goods are represented by graphics, and the numbers in the graphics are used to distinguish the names of the goods (for example, the one displayed in the middle of the graphic is the first product, the one displayed in the middle of the graphic is the second product, and so on). Different shapes of graphics represent different secondary categories of goods. The dotted line frame represents the sequence to be sorted, the solid line frame represents the sorted sequence, and the dotted line frame represents the remaining product sequence from the refined sorted sequence.

[0075] The sequence state S31 is the initial state, the sorted sequence and the to-be-sorted sequence are both empty (not shown in the figure), and all commodities in the refined sorted sequence are stored in the remaining commodity sequence from the refined sorted sequence.

[0076] Sequence state S32, the first product is taken from the head of the remaining product sequence in the finely sorted sequence as the current product. Since the sequence to be sorted is empty, the first product is stored in the end of the sorted sequence. Then, the second product is taken from the head of the remaining product sequence in the finely sorted sequence as the current product. However, the secondary category of the second product is the same as that of the first product, so the second product is stored in the end of the sequence to be sorted. Then, the third product is taken from the head of the remaining product sequence in the finely sorted sequence as the current product. However, the secondary category of the third product is also the same as that of the first product, so the third product is stored in the end of the sequence to be sorted.

[0077] Sequence state S33, since the secondary categories of the second and third products in the to-be-sorted sequence are the same as the first product, the fourth product is taken from the head of the remaining product sequence in the refined sequence as the current product, and since the secondary category of the fourth product is different from that of the first product, the fourth product is stored at the end of the sorted sequence, and the fifth product is taken from the head of the remaining product sequence in the refined sequence as the current product, and since the secondary category of the fifth product is different from that of the first and fourth products, the fifth product is stored at the end of the sorted sequence.

[0078] In sequence state S34, since the second product in the to-be-sorted sequence has a different secondary category from the fourth and fifth products at the end of the sorted sequence, the second product is stored at the end of the sorted sequence, and the third product is at the head of the to-be-sorted sequence.

[0079] Sequence state S35, since the third product in the to-be-sorted sequence has the same secondary category as the second product at the end of the sorted sequence, the sixth product is taken from the head of the remaining product sequence in the finely sorted sequence as the current product, and since the secondary category of the sixth product is different from both the fifth product and the second product at the end of the sorted sequence, the sixth product is stored at the end of the sorted sequence.

[0080] Sequence state S36, since the third product in the to-be-sorted sequence has the same secondary category as the second product at the end of the sorted sequence, the seventh product is taken from the head of the remaining product sequence in the finely sorted sequence as the current product, and since the secondary category of the seventh product is different from both the sixth product and the second product at the end of the sorted sequence, the seventh product is stored at the end of the sorted sequence.

[0081] Sequence state S37: since the third product in the to-be-sorted sequence has different secondary categories from the sixth and seventh products at the end of the sorted sequence, the third product is stored at the end of the sorted sequence, and the to-be-sorted sequence is cleared.

[0082] Sequence state S38, since the to-be-sorted sequence is empty, the eighth product is taken from the head of the remaining product sequence in the refined sorted sequence as the current product, and since the secondary category of the eighth product is different from the seventh product and the third product at the end of the sorted sequence, the eighth product is stored in the end of the sorted sequence, and the ninth product is taken from the head of the remaining product sequence in the refined sorted sequence as the current product, and since the secondary category of the ninth product is different from the third product and the eighth product at the end of the sorted sequence, the ninth product is stored in the end of the sorted sequence, and all products are arranged, and the sorted sequence is used as a new sequence.

[0083] The sequence state S31 is arranged to form the sequence state S38, so that the secondary categories of two adjacent commodities in the generated mixed sequence are different.

[0084] A combination of precise and mixed sorting is adopted, and the sequence algorithm to be sorted is especially broken up according to secondary categories to prevent users from seeing similar products continuously, thereby improving the click-through rate and conversion rate of recommended products on shopping websites.

[0085] Based on a general inventive concept, the present invention also provides a commodity recommendation device.

[0086] Figure 4 is a schematic diagram of a commodity recommendation device provided by an embodiment of the present invention. Figure 4As shown, the product recommendation device includes: a product recall module 41, a feature construction module 42, and a sorting module 43. The product recall module 41 is used to obtain a recalled product set for each user by adopting a multi-way recall method, and the product recall module includes at least two of the multi-way recall including a label recommendation recall unit, a collaborative filtering recommendation recall unit, a popular recommendation recall unit, a user history recall unit, and a vector recall unit; the feature construction module 42 is used for feature construction, and the constructed features include: user browsing features, click features, purchase features, customer layer features, gender features, and differentiated features of products; the sorting module 43 includes a precise sorting unit and a mixed sorting unit. Among them, the label recommendation recall unit is used to perform label recommendation recall on products, the collaborative filtering recommendation recall unit is used to perform collaborative filtering recommendation recall on products, the user history recall unit is used to perform user history recall on products, and the vector recall unit is used to perform vector recall on products. The fine ranking unit uses the LightGBM algorithm model to train and predict the click-through rate of the products in the recalled product set of each user based on the constructed features, and sorts the products in the recalled product set for each user from large to small according to the predicted click-through rate to obtain a fine ranking sequence for each user; the mixed ranking unit breaks up the fine ranking sequence for each user according to the secondary categories to form a mixed ranking sequence for each user, and uses the mixed ranking sequence for each user as the recommended sequence for each user.

[0087] In some embodiments, the popular recommendation recall sorts the popular products in the community according to click-through rate, conversion rate, number of repeat purchases in the community, and online sales, selects multiple products and puts them into the recalled product set, and uses Bayesian smoothing correction for the click-through rate and conversion rate.

[0088] In some embodiments, the constructed features are screened according to the feature importance of the LightGBM algorithm model, and the number of features is reduced by screening the features.

[0089] In some embodiments, the step of breaking up the finely sorted sequence of each user according to the secondary categories to form a mixed sequence further includes: setting a sorted sequence and a sequence to be sorted, the sorted sequence is used to store the products that have been sorted, and the sequence to be sorted is used to store the products that have been taken out of the finely sorted sequence but not yet sorted; looping the following steps until all the products in the finely sorted sequence are stored in the sorted sequence: selecting the current product, taking a product from the head of the sequence to be sorted as the current product, if the sequence to be sorted is empty or all the products in the sequence to be sorted have the same secondary category as the two products at the end of the sorted sequence, then taking a product from the head of the remaining product sequence in the finely sorted sequence as the current product; judging whether the current product and the two products at the end of the sorted sequence have the same secondary category, if they have the same secondary category, then storing the current product at the end of the sequence to be sorted, if they do not have the same secondary category, then storing the current product at the end of the sequence to be sorted; after all products are sorted, use the sorted sequence as the mixed sequence.

[0090] It should be noted that, in this document, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion. The various embodiments in this specification are described in a related manner, and the same and similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0091] The above is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. It should be noted that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A product recommendation method, characterized in that: include: A recalled product set for each user is obtained by a multi-channel recall method, wherein the multi-channel recall includes at least two of tag recommendation recall, collaborative filtering recommendation recall, hot recommendation recall, user history recall, and vector recall; Feature construction, including user browsing features, user click features, user purchase features, user demographic features, user gender features, and product differentiation features; According to the constructed features, the LightGBM algorithm model is used to train and predict the click-through rate of the products in the recalled product set of each user, and the products in the recalled product set for each user are sorted from large to small according to the predicted click-through rate to obtain a refined sorting sequence for each user; The refined sequence for each user is scattered according to the secondary categories to form a mixed sequence for each user, and the mixed sequence for each user is used as the recommended sequence for each user.

2. The method according to claim 1, characterized in that The labels selected for the label recommendation recall include the final category value, customer segment value, brand value, and whether it is a promotion.

3. The method according to claim 1, characterized in that The collaborative filtering recommendation recall calculates the similarity between products through the user's ordering behavior. Among them, A and B represent two commodities respectively.

4. The method according to claim 1, characterized in that: The popular recommendation recall sorts the popular products in the community according to click-through rate, conversion rate, number of repeat purchases in the community, and online sales, selects multiple products and puts them into the recalled product set, and uses Bayesian smoothing correction for the click-through rate and conversion rate.

5. The method according to claim 4, characterized in that Product click-through rate CTR = N / M, N is the number of clicks, M is the exposure, and the Bayesian smoothing correction formula is: in, is the corrected CTR, is the mean, S 2 is the variance, Conversion rate = product sales / (number of clicks to view product details + number of clicks to add to cart).

6. The method according to claim 1, characterized in that The user history recall selects multiple commodities that the user has searched, browsed, and purchased in the history and puts them into the recalled commodity set.

7. The method according to claim 1, characterized in that The vector recall obtains user vectors and product vectors through a word vector generation model, builds a deep semantic matching model, trains the deep semantic matching model and predicts the similarity between the user side and the product side, and finally sorts the predicted similarity scores from large to small, selects multiple products and puts them into the recalled product set. The label during the training of the deep semantic matching model is whether it is clicked.

8. The method according to claim 1, characterized in that In the step of using the LightGBM algorithm model to train and predict the click rate of the products in the recalled product set of each user, the constructed features are selected according to the feature importance of the LightGBM algorithm model, and the LightGBM algorithm model is run in a dynamic partitioned distributed manner.

9. The method according to claim 1, characterized in that: The step of breaking up the refined sequence for each user according to the secondary categories to form a mixed sequence for each user further includes: Setting a sorted column and a to-be-sorted column, wherein the sorted column is used to store the commodities that have been sorted, and the to-be-sorted column is used to store the commodities that have been taken out of the refined sorted column but have not been sorted; The following steps are executed in a loop until all the products in the refined sorting sequence are stored in the sorted sequence: the current product is selected, and a product is taken from the head of the sequence to be sorted as the current product; if the sequence to be sorted is empty or all the products in the sequence to be sorted have the same secondary category as the two products at the end of the sorted sequence, a product is taken from the head of the remaining product sequence in the refined sorting sequence as the current product; Determine whether the current product and the two products at the end of the sorted column have the same secondary category. If they do, store the current product at the end of the column to be sorted. If they do not, store the current product at the end of the column to be sorted. After all the commodities are arranged, the arranged sequence is used as the mixed arrangement sequence.

10. A commodity recommendation device, characterized in that: include: Product recall module: Used to obtain a recalled product set for each user by adopting a multi-channel recall method, wherein the product recall module includes at least two of the multi-channel recall including a tag recommendation recall unit, a collaborative filtering recommendation recall unit, a hot recommendation recall unit, a user history recall unit, and a vector recall unit; The tag recommendation recall unit is used to perform tag recommendation recall on products, the collaborative filtering recommendation recall unit is used to perform collaborative filtering recommendation recall on products, the user history recall unit is used to perform user history recall on products, and the vector recall unit is used to perform vector recall on products; Feature building blocks: Used for feature construction, including: user browsing features, click features, purchase features, customer group features, gender features, and product differentiation features; Sorting module: Including fine arrangement unit and mixed arrangement unit; The refined ranking unit uses the LightGBM algorithm model to train and predict the click-through rate of the products in the recalled product set of each user according to the constructed features, and sorts the products in the recalled product set for each user from large to small according to the predicted click-through rate to obtain a refined ranking sequence for each user; The shuffling unit breaks up the refined sequence for each user according to secondary categories to form a shuffled sequence for each user, and uses the shuffled sequence for each user as a recommended sequence for each user.

Images