An article selection method, device, terminal device, and storage medium

Abstract

An embodiment of the present invention discloses an item selection method, device, terminal device and storage medium. The method includes: obtaining candidate items, generating first feature vectors of each candidate item according to the click sequence of the candidate items, and generating second feature vectors of each candidate item according to the title of the candidate items; determining sample items from the candidate items, and training a selection model according to the first feature vectors and second feature vectors of the sample items; taking the items other than the sample items among the candidate items as items to be selected, and inputting the first feature vectors and second feature vectors of the items to be selected into the trained selection model; determining a target item according to the output item of the trained selection model and the sample item. It can achieve accurate selection of items, which can not only meet the user's needs, but also reduce costs and improve the marketing effect.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of e-commerce data analysis, and in particular, to an item selection method, device, terminal device, and storage medium. Background Art

[0002] By planning themed activities such as hot events and holidays and conducting targeted operations, the profitability and sales volume of e-commerce platforms can be significantly improved. During the targeted operation process, a batch of items highly relevant to the activity theme need to be selected as the item pool for the operation channel.

[0003] In the prior art, the method for selecting highly relevant items is usually to select highly relevant items based on the input selection instruction.

[0004] In the process of implementing the present invention, the inventors found that there are at least the following technical problems in the prior art: users mainly combine the sales volume of items in the same period of previous years and the current theme to input selection instructions. When the selection range is small, it will lead to low item richness and cannot meet user needs; when the selection range is too large, not only will the costs of procurement and sales coordination and source goods preparation increase, but also the operation effect will be affected. Summary of the Invention

[0005] The embodiments of the present invention provide an item selection method, device, terminal device, and storage medium, which can achieve accurate item selection, not only meet user needs, but also reduce costs and improve marketing effects.

[0006] In a first aspect, the embodiments of the present invention provide an item selection method, including:

[0007] Obtain candidate items, generate a first feature vector for each candidate item according to the click sequence of the candidate item, and generate a second feature vector for each candidate item according to the title of the candidate item;

[0008] Determine sample items from the candidate items, and train a selection model according to the first feature vector and the second feature vector of the sample items;

[0009] Use the items other than the sample items among the candidate items as items to be selected, and input the first feature vector and the second feature vector of the items to be selected into the trained selection model;

[0010] Determine target items according to the output items of the trained selection model and the sample items.

[0011] In a second aspect, the embodiments of the present invention provide an item selection device, including:

[0012] A feature generation module, configured to obtain candidate items, generate a first feature vector for each of the candidate items according to the click sequence of the candidate items, and generate a second feature vector for each of the candidate items according to the title of the candidate items;

[0013] A model training module, configured to determine sample items from the candidate items, and train a selection model according to the first feature vector and the second feature vector of the sample items;

[0014] A model application module, configured to use the items other than the sample items among the candidate items as items to be selected, and input the first feature vector and the second feature vector of the items to be selected into the trained selection model;

[0015] A selection module, configured to determine a target item according to the output item of the trained selection model and the sample items.

[0016] In a third aspect, an embodiment of the present invention provides a terminal device, including:

[0017] One or more processors;

[0018] A memory, configured to store one or more programs;

[0019] When the one or more programs are executed by the one or more processors, the one or more processors implement the item selection method as described in any embodiment of the present invention.

[0020] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the item selection method as described in any embodiment of the present invention is implemented.

[0021] An item selection method, device, terminal device and storage medium provided by an embodiment of the present invention obtain candidate items, generate a first feature vector for each candidate item according to the click sequence of the candidate items, and generate a second feature vector for each candidate item according to the title of the candidate items; determine sample items from the candidate items, and train a selection model according to the first feature vector and the second feature vector of the sample items; use the items other than the sample items among the candidate items as items to be selected, and input the first feature vector and the second feature vector of the items to be selected into the trained selection model; determine a target item according to the output item of the trained selection model and the sample items.

[0022] By generating the first feature vector of each candidate item according to the click sequence of the candidate items, it is possible to encode similar items as vectors with a smaller distance based on the user's click behavior on the items in the page, so as to obtain the first feature vector representing the items in the dimension of the user's click behavior. By generating the second feature vector of each candidate item according to the title of the candidate item, it is possible to convert the title of the item itself into a vector, so as to obtain the second feature vector representing the items in the title dimension. By marking sample items from the candidate items, it is possible to achieve multi-target marking in combination with the scenario of selecting items highly relevant to the theme. By training the selection model according to the first feature vector and the second feature vector of the sample items obtained by marking, it is possible to accurately select items highly relevant to the theme for the item to be selected based on the first feature vector and the second feature vector of the item to be selected based on the selection model, which can not only meet the user's needs, but also reduce costs and improve the marketing effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0024] Figure 1 FIG. shows a flowchart of an item selection method provided in Embodiment 1 of the present invention;

[0025] Figure 2 FIG. shows an architecture diagram of an item selection method provided in Embodiment 1 of the present invention;

[0026] Figure 3 FIG. shows a schematic diagram of the steps of generating the first feature vector in an item selection method provided in Embodiment 3 of the present invention;

[0027] Figure 4 FIG. shows a schematic structural diagram of an item selection device provided in Embodiment 6 of the present invention;

[0028] Figure 5 FIG. shows a schematic hardware structure diagram of a terminal device provided in Embodiment 7 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will describe the technical solutions of the present invention clearly and completely with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention. In the following embodiments, optional features and examples are provided simultaneously in each embodiment. The features described in the embodiments can be combined to form multiple alternative solutions. Each numbered embodiment should not be regarded as only one technical solution.

[0030] Embodiment 1

[0031] Figure 1 The flowchart of an item selection method provided in Embodiment 1 of the present invention is shown. The item selection method provided in the embodiments of the present invention is applicable to the situation of item selection, such as the situation of selecting items with a high degree of relevance to themes such as hot activities and holidays. This method can be executed by an item selection device, which is implemented in software and / or hardware and can be configured in a terminal device, such as configured in the server of an e-commerce platform.

[0032] As Figure 1 shown, the item selection method provided in the embodiments of the present invention includes the following steps:

[0033] S110. Obtain candidate items, generate a first feature vector for each candidate item according to the click sequence of the candidate items, and generate a second feature vector for each candidate item according to the title of the candidate items.

[0034] In the embodiments of the present invention, the candidate items can be the items listed on the e-commerce platform or the items preliminarily screened from the listed items and related to the theme. The click sequence of the candidate items can be considered as the sequence of click behaviors of each user terminal of the e-commerce platform on the candidate items in the page. The title of the candidate items can be considered as the title set by each merchant terminal of the e-commerce platform for the listed items.

[0035] The server of the e-commerce platform can store the relevant data of the listed items in a database, such as stored in the Hadoop Distributed File System (HDFS). Among them, the relevant data of the items can include, but are not limited to, the basic attribute data of the items (such as the title, specifications, parameters, etc. of the items) and the interaction data of the items (such as the data of interaction behaviors such as search, click, favorite, purchase, etc. of the user terminal on the items).

[0036] The server can extract the listed items from the database as candidate items through data processing methods, such as the data processing method mainly based on Hive. Or it can also preliminarily screen the items related to the theme based on traditional mathematical algorithms or deep learning models according to the relevant data of the listed items (such as search volume, sales volume, etc.) to obtain candidate items. By preliminarily screening the listed items, the computing resources consumed in the subsequent selection process can be saved to a certain extent.

[0037] After obtaining the candidate items, the server can also extract the click sequence and title corresponding to the candidate items from the database through data processing methods. Furthermore, based on the data vector conversion method, the click sequence can be converted into a first feature vector, and the title can be converted into a second feature vector. Among them, the data vector conversion method can include, but is not limited to, methods of converting data into vectors by pre-trained models (Pre-trained Models, PTMs) such as models based on shallow word embeddings (such as word2vec and GloVe), models based on pre-trained encoders (such as BERT and ELMo), and models based on supervised learning (such as CoVe).

[0038] By generating the first feature vector of each candidate item according to the click sequence of the candidate item, it is possible to encode similar items as vectors with a smaller distance based on the user's click behavior on the items in the page, so as to obtain the first feature vector representing the item in the user click behavior dimension. By generating the second feature vector of each candidate item according to the title of the candidate item, it is possible to convert the title of the item itself into a vector, so as to obtain the second feature vector representing the item in the title dimension.

[0039] S120. Determine sample items from the candidate items, and train the selection model according to the first feature vector and the second feature vector of the sample items.

[0040] Among them, determining sample items from the candidate items can be considered as marking the sample items among the candidate items. In the process of marking sample items, the marking logic is related to the learning task of the model. For example, if the learning task of the model is to predict items with high sales volume, the marking logic is to mark items with high sales volume as sample items; another example is that if the learning task of the model is to predict items with high search volume, the marking logic is to mark items with high search volume as sample items.

[0041] In this embodiment, the learning task of the selection model is to select items that are highly relevant to the theme of the activity. Since items that are highly relevant to the theme can be defined in multiple dimensions, the learning task can be broken down into multiple subtasks. Among them, the multiple subtasks may include at least two of the following: predicting items with high sales, predicting items with high search volumes, predicting items with high collection volumes, predicting items with high purchase volumes, and predicting items with high return visits. In addition, the subtasks may also include other tasks that can be broken down from the task of selecting items that are highly relevant to the theme, which are not exhaustively listed here. Accordingly, the labeling logic of the selection model can be considered to be the labeling logic corresponding to multiple subtasks, which can be called hot spot labeling logic. The sample items marked based on the hot spot labeling logic can be considered to be items that meet the requirements of high relevance to the theme in multiple dimensions.

[0042] Among them, the selected model can be a deep learning model that can realize multi-task learning, for example, it can be a multi-task learning task relationship modeling (Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts, MMoE) model, or it can be a deep factorization machine (Deep factorization machine, DeepFM) model, which are not listed here exhaustively.

[0043] By training the selection model according to the first feature vector and the second feature vector of the sample item, the selection model can learn the logical relationship between the first feature vector, the second feature vector, and the relevance to the topic. Thus, items highly relevant to the topic can be selected based on the trained selection model.

[0044] Compared with the traditional single-target deep learning algorithm, this embodiment combines the scene of selecting items that are highly relevant to the topic for multi-target labeling, which can significantly improve the prediction accuracy of the trained selection model.

[0045] S130: Items other than the sample item among the candidate items are taken as items to be selected, and the first feature vector and the second feature vector of the items to be selected are input into the trained selection model.

[0046] During the process of marking samples, a small portion of candidate items can be marked as sample items, and the unmarked items among the candidate items can be regarded as items to be selected. After training a selection model based on the first feature vector and the second feature vector of the sample items, the trained selection model can be used to select the items to be selected. Specifically, the first feature vector and the second feature vector of the items to be selected are input into the trained selection model, so that the selection model can obtain the degree of relevance of each item to be selected to the theme. This degree of relevance can be represented by a probability value in the range of (0, 1), and the larger the probability value, the higher the relevance to the theme. Furthermore, the selection model can determine the candidate items with probability values greater than a first preset value as the selected items for output, where the first preset value can be set according to empirical values or experimental values, such as 0.8, 0.9, etc.

[0047] S140. Determine the target item according to the output items of the trained selection model and the sample items.

[0048] After the selection model outputs items highly relevant to the theme, the output items and the sample items can be used as the selected items. In some implementation manners, the first feature vector and the second feature vector of all candidate items including the sample items can also be input into the selection model, and the output items of the selection model can be used as the selected items. Compared with the solution where the selection model only selects the items to be selected, the computational amount is relatively larger, but the accurate selection of items highly relevant to the theme can also be achieved. After determining the selected items, an item pool for a recommended channel or a promotion activity page in the theme activity can be constructed according to the selection to support services such as recommendation and search.

[0049] In this embodiment, for the scenario of selecting items highly relevant to the theme, by determining the feature vectors of each candidate item in the dimensions of user click behavior and title, marking sample items according to the hot tagging logic, training a selection model based on the feature vectors of the sample items, and performing selection according to the selection model, the accurate selection of items highly relevant to the theme can be realized. Accurately selecting items can lay a foundation for constructing an item pool for an operation channel, which can not only ensure that the item categories meet the user needs, but also reduce the costs consumed by procurement and sales coordination and source goods preparation. At the same time, it can improve the conversion rate from visit to purchase to a certain extent and has a positive impact on increasing the average sales per visitor, thereby improving the operation effect of the theme activity.

[0050] In some alternative embodiments, the item selection method further includes: determining a third feature vector of a candidate item based on at least one of the following data of the candidate item: the price range it belongs to, the favorable review rate, and the sales volume of items in the same category; correspondingly, training the selection model based on the first feature vector and the second feature vector of the sample item includes: training the selection model based on the first feature vector, the second feature vector, and the third feature vector of the sample item; inputting the first feature vector and the second feature vector of the item to be selected into the trained selection model includes: inputting the first feature vector, the second feature vector, and the third feature vector of the item to be selected into the trained selection model.

[0051] In these alternative embodiments, in addition to determining the feature vectors of each candidate item in the user click behavior dimension and the title dimension, the feature vectors of the candidate item can also be determined in dimensions such as the price range it belongs to, the favorable review rate, and the sales volume of items in the same category, that is, determining the third feature vector. Correspondingly, during the training process of the selection model, the selection model can be trained using the first feature vector, the second feature vector, and the third feature vector of the sample item; during the prediction process of the selection model, the selection model can select items highly relevant to the theme based on the first feature vector, the second feature vector, and the third feature vector of the item to be selected. By introducing the third feature vector, the selection accuracy of the item can be improved to a certain extent.

[0052] Exemplarily, Figure 2 shows an architecture diagram of an item selection method provided in Embodiment 1 of the present invention. Refer to Figure 2 , after obtaining the candidate item, the interaction data (such as click sequence, search, purchase, etc. data) and basic attribute data (such as title, price, favorable review rate, etc. data) of the candidate item can be extracted; generating a first feature vector according to the click sequence, generating a second feature vector according to the title, generating a third feature vector according to other data in the basic attribute data, and each feature vector can be used as the feature data of the item; marking the sample item according to the hot topic marking logic and the interaction data, and the feature data of the sample item can be extracted from the feature data; training the selection model according to the feature data of the sample item; based on the trained selection model, outputting an item according to the feature data of the item to be selected, and using the output item and the sample item as the target item.

[0053] An item selection method provided by an embodiment of the present invention can generate a first feature vector for each candidate item according to the click sequence of the candidate items, and can encode similar items into vectors with smaller distances based on the user's click behavior on the items in the page, so as to obtain a first feature vector representing the items in the dimension of the user's click behavior. By generating a second feature vector for each candidate item according to the title of the candidate item, the title of the item itself can be converted into a vector, so as to obtain a second feature vector representing the items in the title dimension. By marking sample items from the candidate items according to the hot marking logic, multi-target marking can be realized by combining the scenario of selecting items highly relevant to the theme. By training the selection model according to the first feature vector and the second feature vector of the marked sample items, based on the selection model, according to the first feature vector and the second feature vector of the item to be selected, accurate selection of items highly relevant to the theme for the item to be selected can be realized, which can not only meet the user's needs, but also reduce costs and improve the marketing effect.

[0054] Embodiment 2

[0055] On the basis of the above embodiment, this embodiment describes in detail the steps of obtaining candidate items.

[0056] In this embodiment, obtaining candidate items may include:

[0057] Determine hot items according to the items clicked in the search results corresponding to the search terms containing hot words within the first preset time period;

[0058] For each item category, determine the first ratio of the number of hot items included to the total number of categories, and / or determine the second ratio of the sales volume of the hot items included to the total number of categories;

[0059] According to the first ratio and / or the second ratio, screen out candidate categories from each item category, and use the items under the candidate categories as candidate items.

[0060] In this embodiment, during the process of determining hot items, it may include: within a first preset period before the start time of the themed event, the server records the search terms of each client and filters out target search terms that contain preset hot terms; after pushing the search results corresponding to the target search terms to the client, it may record the items among these pushed items that are clicked by the client, which can be understood as the accessed items; among them, the first or the first N clicked items can be recorded; the server can count the number of times each listed item is clicked after being pushed as a pushed item, and take the items with the number of clicks greater than a second preset value as hot items. Among them, the first preset period and the second preset value can be set according to empirical values or experimental values. For example, the first preset period can be three weeks, one month, etc., and the second preset value can be one million times, two million times, etc.

[0061] In this embodiment, during the process of screening candidate categories, it is possible to consider from different aspects whether a category is the current hot category. And the first aspect can be the first ratio of the number of hot items included in a certain item category to the total number of items in that category; the second aspect can be the second ratio of the sales volume of the hot items included in a certain item category to the total number of items in that category. Among them, the higher the first ratio, the more hot items can be considered to be covered under that item category, and the higher the probability that that item category is a hot category; the higher the second ratio, the better the average sales volume of that item category can be considered, and the higher the probability that that item category is a hot category. In addition, it is also possible to consider from other aspects whether a category is the current hot category, which will not be enumerated here.

[0062] Correspondingly, the server can screen candidate categories according to the first ratio and / or the second ratio. For example, the first M items ranked by the first ratio can be used as candidate categories; or the first M items ranked by the second ratio can be used as candidate categories; or the M items among the first Q items ranked by the first ratio and also among the first Q items ranked by the second ratio can be used as candidate items. Among them, increasing the value of N can improve the richness of the selected items, but also correspondingly increase the calculation amount of the subsequent selection. Q and M are usually set to values that balance the richness and the calculation amount according to experimental values. For example, Q can be 100 and M can be 50, which are not specifically limited here.

[0063] Exemplarily, Table 1 can show the situation of the first ratio and the second ratio of some categories within 1 month before the Mother's Day event. Referring to Table 1, the item categories can be the categories preset at a certain level on the e-commerce platform, such as the third-level categories. Table 1 can include the third-level category names, the total number of items under the category, the first ratio, the sales volume of the hot items, and the second ratio. Among them, candidate items can be determined according to the first ratio and / or the second ratio.

[0064] Table 1

[0065]

[0066] In this embodiment, the items on the shelves can be filtered to obtain candidate categories, and the items under the candidate categories are used as candidate items, which can not only improve the selection accuracy of the items, but also reduce the consumption of computing resources in the selection process.

[0067] Based on the above embodiment, the embodiment of the present invention describes in detail the steps of obtaining candidate items. By using a first ratio representing the proportion of hot items in a category and a second ratio representing the sales volume of hot items, it is possible to screen out categories with a relatively high occupancy rate of hot items from among numerous item categories. Furthermore, by using the items under the screened-out categories as candidate items, it is possible to not only improve the selection accuracy of the items, but also reduce the consumption of computing resources in the selection process. In addition, the item selection method proposed in the embodiment of the present invention and the above embodiment belong to the same inventive concept. Technical details not described in detail in this embodiment can be referred to in the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0068] Embodiment Three

[0069] Based on the above embodiment, the steps of generating the first feature vector are described in detail. Exemplarily, Figure 3 shows a schematic diagram of the steps of generating the first feature vector in an item selection method provided in Embodiment Three of the present invention. Figure 3 It may include (a)-(d), and may respectively illustrate the steps of generating the first feature vector, as follows:

[0070] Figure 3 In (a), it shows obtaining the click sequences of each client on the candidate items within a second preset time period. Referring to (a), the server can obtain the click sequences of clients U1-U3 on candidate items A-F within the second preset time period, and can sequentially connect the candidate items in the click sequences to obtain the actual click stream of the items corresponding to each client. The second preset time period can be set according to empirical values or experimental values, for example, it can be three months.

[0071] Figure 3(b) illustrates that the item knowledge graph is constructed according to each click sequence, which can be specifically: constructing the item knowledge graph according to the actual click stream of each item. Referring to (b), each node in the item knowledge graph can represent each candidate item AF, and the connection between each node has directionality and weight. Among them, when the interval between click behaviors in the item click sequence is less than the first preset time, that is, when there is no dotted line between the candidate items in (a), a connection between the candidate items corresponding to the two click behaviors can be constructed, and the direction of the connection is from the candidate item at the front of the time sequence to the candidate item at the back of the time sequence. Among them, the weight of the connection can be considered as the number of co-occurrences of each directional connection counted by the server, and the greater the number of co-lines, the greater the weight of the connection can be considered. Among them, the first preset time can be set according to an empirical value or an experimental value, for example, it can be half an hour.

[0072] Figure 3 (c) illustrates the determination of sample click streams based on the item knowledge graph, which can be specifically as follows: determining the starting node from the item knowledge graph, and performing random walks according to the weights to obtain the sample click stream of the item. Referring to (c), the starting node can be any candidate item such as A, B, C, D or E. According to the weighted random walk, it can be considered that the greater the weight of the connection between the two candidate items, the greater the probability of walking through the two candidate items in succession. For example, in (c), B walked from A 3 times and B walked from E 5 times, so it is highly likely that the weight of the edge from B to E in (b) is higher than the weight of the edge from B to A. The window value of the random walk (the value of the item node in the sample click stream of the item) can be a fixed value or a non-fixed value. For example, the window value of the sample click stream of the item in (c) can be 4 or 5.

[0073] Figure 3 (d) shows that based on the first vector model, the first feature vector of each candidate item is generated according to the sample click stream. See (d) for a schematic diagram of the structure of the first vector model. The sample click stream of each item can be used as the input of the model. The candidate items that appear in the same window are determined as positive samples through the hidden layer, and those that do not appear in the same window are negative samples. The output of the first feature vector of each candidate item is achieved by obtaining the weight of the implicit function.

[0074] In this embodiment, the first vector model may belong to the models of PTMs. For example, it may be a Skip-Gram model based on Graph-Embedding. By generating the first feature vector through the Skip-Gram model, the connection of the click order between candidate items can be transformed into a vector representation form. The higher the vector similarity, the closer the relationship between the candidate items can be considered. Among them, the first feature vector may be, for example, a 32-dimensional embedding feature vector. In this embodiment, after constructing the item knowledge graph based on the actual click stream of items, Graph-Embedding is used to encode the feature vectors of candidate items, realizing the application of Graph-Embedding in the scenario of selecting items highly relevant to the theme.

[0075] Based on the above embodiment, the embodiment of the present invention describes in detail the step of generating the first feature vector. By constructing an item knowledge graph according to the item click sequences of candidate items on each client, a weighted random walk graph is used to obtain the sample click stream of items; the first vector model is used to encode similar items into vectors with smaller distances according to the front-back relationship of items in the sample click stream of items, so as to generate the first feature vector that can represent candidate items in the dimension of user click behavior. In addition, the item selection method proposed in the embodiment of the present invention and the above embodiment belongs to the same inventive concept. Technical details not described in detail in this embodiment can be referred to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0076] Embodiment 4

[0077] Based on the above embodiment, the embodiment of the present invention describes in detail the step of generating the second feature vector. In this embodiment, generating the second feature vector of each candidate item according to the title of the candidate item includes:

[0078] Preprocessing the titles of the candidate items to obtain reconstructed titles; based on the second vector model, generating the second feature vectors of the candidate items according to the reconstructed titles. Among them, the preprocessing may include at least one of the following processes: filtering non-Chinese characters, word segmentation, filtering word segmentation based on word frequency, and filtering word segmentation based on part of speech.

[0079] In this embodiment, filtering non-Chinese characters from the item title helps with word segmentation. The title after filtering non-Chinese characters can be segmented by a word segmentation tool, and part-of-speech tagging can also be performed. The results after word segmentation can be filtered according to word frequency and part-of-speech; among them, the word segmentation with a word frequency less than the third preset value can be filtered. Through experimental verification, filtering words with a word frequency less than 10 times can significantly improve the title reconstruction speed and will not affect the calculation of the subsequent second feature vector; among them, the expected part-of-speech to be filtered can be preset in combination with the theme of the current specific activity, and the title can be filtered to a certain extent according to the preset part-of-speech, so that the second feature vector can better represent the degree of association with the theme.

[0080] Among them, the preprocessed title can be used as the reconstructed title, and the vectors of each word segmentation in the reconstructed title can be obtained according to the second vector model. The second feature vector of the reconstructed title can be obtained according to the vectors of each word segmentation. For example, the vectors of each word segmentation can be averaged to obtain the second feature vector of the reconstructed title.

[0081] In this embodiment, the second vector model can belong to the PTMs model, for example, it can be a BERT model. By generating the second feature vector through the BERT model, the candidate item title can be transformed into a vector representation form. The higher the vector similarity, the more similar the candidate items are considered. Among them, the second feature vector can be, for example, a 768-dimensional embedding feature vector. In this embodiment, by preprocessing the item title, the BERT model can be well supported for vector transformation at the phrase level in the item title scenario.

[0082] On the basis of the above embodiment, the embodiment of the present invention describes in detail the steps of generating the second feature vector. By preprocessing the title of the item, the second feature vector generated based on the second vector model can more accurately represent the item title. In addition, the item selection method proposed in the embodiment of the present invention and the above embodiment belongs to the same inventive concept. The technical details not described in detail in this embodiment can be referred to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0083] Embodiment Five

[0084] On the basis of the above embodiment, this embodiment describes in detail the steps of determining the sample item. In this embodiment, determining the sample item from the candidate items includes: determining the first sample item from the candidate items, where the number of clicks of the first sample item is higher than the first threshold; determining the second sample item from the candidate items, where the sales volume of the second sample item during the historical hot period is higher than the second threshold.

[0085] In this embodiment, the learning task of the selection model can be decomposed into two subtasks, namely, predicting items with high sales volume and predicting items with high search volume. Correspondingly, the labeling logic of the selection model can be considered as the labeling logic of labeling items with high sales volume and high search volume as sample items, which can be called the hot-spot labeling logic.

[0086] Among them, the historical hot-spot period can refer to the third preset period before the theme activities corresponding to the same hot-spot activities or holidays in the historical year. Among them, the first threshold and the second threshold can be set according to empirical values or experimental values. Usually, it is difficult to collect the first sample items labeled based on the number of clicks of the current theme, and it is relatively easy to obtain the second sample items with high sales volume during the historical hot-spot period. Therefore, the number of the first sample items can be less than the number of the second sample items. For example, the number of the first sample items can be 50,000, and the number of the second sample items can be 200,000. Among them, the first threshold is usually greater than the second preset value used to determine hot-spot items during the candidate item acquisition process. For example, if the second preset value is 2 million, the first threshold can be set to 3 million.

[0087] Among them, labeling based on the logic that the number of clicks is higher than the first threshold can be considered that the labeled first sample items are the currently highly searched items; labeling based on the logic that the sales volume during the historical hot-spot period is higher than the second threshold can be considered that the labeled second sample items are the historical high-order items. The sample items labeled based on the hot-spot labeling logic can be considered as items that meet the requirements of high relevance to the theme.

[0088] In addition, on the basis of the above embodiment, this embodiment also describes in detail the training steps of the selection model. In this embodiment, training the selection model according to the first feature vector and the second feature vector of the sample items may include:

[0089] Dividing the sample items into training samples and test samples; initially training the selection model using the first feature vector and the second feature vector of the training samples, and the training weight of the first sample items in the training samples is higher than that of the second sample items; testing the initially trained selection model using the first feature vector and the second feature vector of the test samples, and determining that the training of the selection model is completed when the test passes.

[0090] In this embodiment, the selection model is described by taking the MMoE model as an example. Among them, by learning the connections and differences between different tasks, the MMoE model can improve the learning efficiency and quality of each task. And the framework of multi-task learning can adopt the structure of a shared bottom network, and the bottom hidden layer network can be shared among different tasks. Based on different tasks, corresponding tower networks can be connected above the bottom network respectively.

[0091] Among them, the MoE model can be expressed as: Among them, f are n expert networks included in the bottom network, and each expert network can be regarded as a neural network; g is a gating network for combining the results of the expert networks. Specifically, g generates a probability distribution of the n expert networks, and the final output is the weighted sum of all the expert networks.

[0092] Among them, the MMoE model takes the MoE model as a basic building block and stacks multiple MoE model structures in a large network. For example, an MoE layer can receive the output of the previous MoE layer as input and use its output as the input for the next layer. Compared with the shared-bottom structure, the MMoE model can capture the differences in tasks without significantly increasing the requirements for model parameters. Its core idea is to replace the function f in the shared-bottom network with an MoE layer.

[0093] In these further embodiments, the sample items can be divided into training samples and test samples according to a certain ratio. For example, the ratio of training samples to test samples is 8:2. During the training process of the MMoE model using the training samples, the training weight of the first sample item is higher than that of the second sample item, that is, the training is mainly based on the first sample item and supplemented by the second sample item. By learning mainly from the sample items with high searches, the selected items can be made more in line with the current theme; by learning from the sample items with high historical orders as a supplement, the learning effect of the model can be further improved, compensating for the problem of low accuracy caused by training the model based on the first sample item.

[0094] It has been experimentally proven that taking the Mother's Day hot topic as an example, the number of candidate items screened is 2 million, among which the number of positive samples marked based on the high-search marking logic is 50,000, and the number of positive samples marked based on the high-order marking logic is 200,000. Setting the ratio of training samples to test samples in the sample items as 8:2, the experimental results are as follows: for the model trained using the first feature vector, the second feature vector, and the third feature vector, the value of the model evaluation index AUC (Area Under the Curve) is 0.9316; for the model trained only using the third feature vector, the value of the model evaluation index AUC (Area Under the Curve) is 0.4704.

[0095] Exemplarily, Table 2 shows the correlation between the probability intervals predicted by the selected model in the above experiment and the degree of relevance to the theme. Referring to Table 2, the larger the probability value in the predicted probability interval, the higher the indicators such as the proportion of ordered items and the order placement probability of the corresponding items. The experimental results prove that the selection method provided in this embodiment is feasible and can achieve accurate selection of items highly relevant to the theme.

[0096] Table 2

[0097] Predicted probability value interval Covered item Proportion of ordered items Ordering probability % (0,0.1) 351907 0.0820 0.44 (0.1,0.2) 15892 0.2468 2.12 (0.2,0.3) 7195 0.3245 4.29 (0.3,0.4) 4327 0.3899 7.52 (0.4,0.5) 2988 0.4495 7.85 (0.5,0.7) 3725 0.5278 15.01 (0.7,1) 2643 0.7068 97.57

[0098] Based on the above embodiments, the present invention embodiment describes in detail the steps of labeling sample items and the training of the selection model. By sampling items with a high degree of relevance to the theme according to the multi-objective labeling logic of high search and high order placement, the selection model can learn multiple tasks of selecting high search and high order placement items. By mainly learning with sample items of high search during the model learning and training process, the selected items can be made more in line with the current theme; by supplementarily learning with sample items of historical high order placement, the learning effect of the model can be further improved.

[0099] In addition, the item selection method proposed in the present invention embodiment and the above embodiments belongs to the same inventive concept. For technical details not described in detail in this embodiment, reference can be made to the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0100] Embodiment Six

[0101] Figure 4 Fig. shows a schematic structural diagram of an item selection device provided in Embodiment Six of the present invention. The present invention embodiment is applicable to the situation of item selection, such as the situation of selecting items with a high degree of relevance to themes such as hot activities and holidays. This method can be executed by an item selection device, which is implemented in a software and / or hardware manner, preferably configured in a terminal device, such as a server of an e-commerce platform and other terminal devices.

[0102] As Figure 4 shown, the item selection device in the present invention embodiment includes:

[0103] A feature generation module 410, configured to obtain candidate items, generate a first feature vector for each candidate item according to the click sequence of the candidate item, and generate a second feature vector for each candidate item according to the title of the candidate item;

[0104] A model training module 420, configured to determine sample items from the candidate items, and train a selection model according to the first feature vector and the second feature vector of the sample items;

[0105] The model application module 430 is configured to use the items other than the sample item among the candidate items as the items to be selected, and input the first feature vector and the second feature vector of the items to be selected into the trained selection model;

[0106] The selection module 440 is configured to determine the target item according to the output item of the trained selection model and the sample item.

[0107] In some alternative embodiments, the feature generation module may include:

[0108] The category screening unit is configured to:

[0109] Determine the hot items according to the clicked items in the search results corresponding to the search terms including the hot words within the first preset time period;

[0110] For each item category, determine the first ratio of the number of hot items included to the total number of items in the category, and / or determine the second ratio of the sales volume of the hot items included to the total number of items in the category;

[0111] Screen out the candidate categories from each item category according to the first ratio and / or the second ratio, and use the items under the candidate categories as the candidate items.

[0112] In some alternative embodiments, the feature generation module may include:

[0113] The first feature vector generation unit is configured to:

[0114] Obtain the click sequences of each client for the candidate items within the second preset time period;

[0115] Construct an item knowledge graph according to each click sequence, and determine a sample click stream based on the item knowledge graph;

[0116] Generate the first feature vector of each candidate item based on the first vector model according to the sample click stream.

[0117] In some alternative embodiments, the feature generation module may include:

[0118] The second feature vector generation unit is configured to:

[0119] Preprocess the title of each candidate item to obtain a reconstructed title;

[0120] Generate the second feature vector of each candidate item based on the second vector model according to each reconstructed title.

[0121] In some alternative embodiments, the model training module may include:

[0122] The multi-objective labeling unit is configured to:

[0123] Determine a first sample item from the candidate items, where the number of clicks on the first sample item is higher than a first threshold;

[0124] Determine a second sample item from the candidate items, where the sales volume of the second sample item during the historical hot period is higher than a second threshold.

[0125] In some alternative embodiments, the model training module may include:

[0126] A training unit for:

[0127] Divide the sample items into training samples and test samples;

[0128] Initially train the selection model using the first feature vector and the second feature vector of the training samples, and the training weight of the first sample item in the training samples is higher than that of the second sample item;

[0129] Test the selection model that has completed the initial training using the first feature vector and the second feature vector of the test samples, and determine that the training of the selection model is completed when the test passes.

[0130] In some alternative embodiments, the feature generation module may further include:

[0131] A third feature vector generation unit for:

[0132] Determine the third feature vector of the candidate item according to at least one of the following data of the candidate item: the price range it belongs to, the favorable comment rate, and the sales volume of items in the same category;

[0133] Correspondingly, training the selection model according to the first feature vector and the second feature vector of the sample item includes: training the selection model according to the first feature vector, the second feature vector, and the third feature vector of the sample item;

[0134] Inputting the first feature vector and the second feature vector of the item to be selected into the trained selection model includes: inputting the first feature vector, the second feature vector, and the third feature vector of the item to be selected into the trained selection model.

[0135] The item selection device provided by the embodiments of the present invention belongs to the same inventive concept as the item selection method provided by the above embodiments. The technical details not described in detail in the embodiments of the present invention can be referred to the above embodiments, and the embodiments of the present invention have the same beneficial effects as the above embodiments.

[0136] Embodiment VII

[0137] Figure 5FIG. 0 shows a schematic hardware structure of a terminal device provided in Embodiment 7 of the present invention. The terminal device in the embodiments of the present invention may include, but is not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Tablet Computers), PMPs (Portable Multimedia Players), in-vehicle terminals (such as in-vehicle navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 5 The terminal device shown is only an example and should not impose any limitation on the functions and usage scope of the embodiments of the present invention.

[0138] As Figure 5 shown, the terminal device 500 may include a processing device (such as a central processing unit, a graphics processing unit, etc.) 501, which may perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 502 or a program loaded from a storage device 508 into a random access memory (RAM) 503. In the RAM 503, various programs and data required for the operation of the terminal device 500 are also stored. The processing device 501, the ROM 502, and the RAM 503 are connected to each other through a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0139] Generally, the following devices may be connected to the I / O interface 505: an input device 506 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 507 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 508 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 509. The communication device 509 may allow the terminal device 500 to communicate with other devices wirelessly or wiredly to exchange data. Although Figure 5 FIG. shows the terminal device 500 having various devices, it should be understood that it is not required to implement or include all the shown devices. More or fewer devices may be alternatively implemented or included.

[0140] Particularly, according to an embodiment of the present invention, the process described above with reference to the flowchart may be implemented as a computer software program. For example, an embodiment of the present invention includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes program codes for performing the method shown in the flowchart. In such an embodiment, the computer program may be downloaded and installed from a network through the communication device 509, or installed from the storage device 508, or installed from the ROM 502. When the computer program is executed by the processing device 501, the above-mentioned functions defined in the item selection method provided in the embodiments of the present invention are executed.

[0141] The terminal provided by an embodiment of the present invention and the article selection method provided by the above embodiment belong to the same inventive concept. For technical details not described in detail in the embodiment of the present invention, reference may be made to the above embodiment, and the embodiment of the present invention has the same beneficial effects as the above embodiment.

[0142] Embodiment Eight

[0143] An embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the article selection method provided by the above embodiment.

[0144] It should be noted that the computer-readable storage medium in the embodiment of the present invention may be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples of the computer-readable storage medium may include, but are not limited to: an electrical connection having 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 a flash memory (FLASH), 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 embodiment of the present invention, the computer-readable storage medium may be any tangible medium that contains or stores a program, and the program can be used by or in combination with an instruction execution system, apparatus, or device. In the embodiment of the present invention, the computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which computer-readable program code is carried. Such a propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The computer-readable signal medium may also be any computer-readable medium other than the computer-readable storage medium, and the computer-readable signal medium can send, propagate, or transmit a program for use by or in combination with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted by any appropriate medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0145] In some embodiments, the client and the server can communicate using any currently known or future-developed network protocol such as HTTP (HyperText Transfer Protocol), and can be interconnected with digital data communication in any form or medium (e.g., a communication network). Examples of communication networks include local area networks ("LANs"), wide area networks ("WANs"), the Internet (e.g., the Internet), and end-to-end networks (e.g., ad hoc end-to-end networks), as well as any currently known or future-developed networks.

[0146] The above computer-readable storage medium can be included in the above terminal device or can exist separately without being assembled into the terminal device.

[0147] The above terminal device stores and carries one or more programs. When the one or more programs are executed by the terminal device, the terminal device is caused to:

[0148] Obtain candidate items, generate a first feature vector for each candidate item according to the click sequence of the candidate items, and generate a second feature vector for each candidate item according to the title of the candidate items; determine sample items from the candidate items, and train a selection model according to the first feature vector and the second feature vector of the sample items; use the items other than the sample items among the candidate items as items to be selected, and input the first feature vector and the second feature vector of the items to be selected into the trained selection model; determine the target item according to the output item and the sample item of the trained selection model.

[0149] Computer program code for performing the operations of the present invention can be written in one or more programming languages or combinations thereof. The above programming languages include object-oriented programming languages such as Java, Smalltalk, C++, and also include conventional procedural programming languages such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, executed as an independent software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (e.g., by using an Internet service provider to connect through the Internet).

[0150] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagram may represent a module, a segment of a program, or a part of code that contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that marked in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, as well as combinations of blocks in the block diagram and / or flowchart, may be implemented by a dedicated hardware-based system that performs the specified functions or operations, or may be implemented by a combination of dedicated hardware and computer instructions.

[0151] The units involved in the embodiments of the present invention can be implemented in software or in hardware. Among them, the name of the unit does not constitute a limitation to the unit itself in some cases.

[0152] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that can be used include: Field Programmable Gate Arrays (FPGAs), Application Specific Integrated Circuits (ASICs), Application Specific Standard Products (ASSPs), Systems on Chip (SOCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0153] Note that the above is only the preferred embodiment of the present invention and the technical principles applied. Those skilled in the art will understand that the present invention is not limited to the specific embodiments here, and various obvious changes, re-adjustments, and substitutions can be made by those skilled in the art without departing from the protection scope of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments. Without departing from the concept of the present invention, more other equivalent embodiments can be included, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. An article selection method, characterized in that, Including: Obtain candidate items, generate first feature vectors of each candidate item according to the click sequence of the candidate item, and generate second feature vectors of each candidate item according to the title of the candidate item; Determine sample items from the candidate items, and train a selection model according to the first feature vectors and second feature vectors of the sample items; Take the items other than the sample items among the candidate items as items to be selected, and input the first feature vectors and second feature vectors of the items to be selected into the trained selection model; Determine the target item according to the output item of the trained selection model and the sample item; The obtaining of the candidate items includes: Determine hot items according to the items clicked in the search results corresponding to the search terms containing hot words within the first preset time period; For each item category, determine the first ratio of the number of hot items included to the total number of categories, and / or determine the second ratio of the sales volume of the hot items included to the total number of categories; According to the first ratio and / or the second ratio, screen out candidate categories from each item category, and take the items under the candidate categories as candidate items; Among them, the determining of the sample items from the candidate items includes: Determine the first sample item from the candidate items, where the number of clicks of the first sample item is higher than the first threshold; Determine the second sample item from the candidate items, where the sales volume of the second sample item is higher than the second threshold during the historical hot period; Among them, the number of the first sample items is less than the number of the second sample items, and the first threshold is greater than the second threshold.

2. The method according to claim 1, wherein The generating of the first feature vectors of each candidate item according to the click sequence of the candidate item includes: Obtain the click sequences of each candidate item on each client within the second preset time period; Construct an item knowledge graph according to each click sequence, and determine a sample click stream based on the item knowledge graph; Based on the first vector model, generate the first feature vectors of each candidate item according to the sample click stream.

3. The method according to claim 1, characterized in that, The generating of the second feature vectors of each candidate item according to the title of the candidate item includes: Preprocess the title of each candidate item to obtain a reconstructed title; Based on the second vector model, generate the second feature vectors of each candidate item according to each reconstructed title.

4. The method according to claim 1, characterized in that, The training of the selection model according to the first feature vectors and second feature vectors of the sample items includes: Divide the sample items into training samples and test samples; Use the first feature vectors and second feature vectors of the training samples to perform initial training on the selection model, and the training weight of the first sample item in the training samples is higher than that of the second sample item; Use the first feature vectors and second feature vectors of the test samples to test the selection model after the initial training is completed, and determine that the selection model is trained when the test passes.

5. The method according to claim 1, characterized in that, It also includes: Determine the third feature vector of the candidate item according to at least one of the following data of the candidate item: the price range it belongs to, the positive review rate, and the sales volume of items in the same category; Correspondingly, training the selection model according to the first feature vector and the second feature vector of the sample item includes: training the selection model according to the first feature vector, the second feature vector, and the third feature vector of the sample item; Inputting the first feature vector and the second feature vector of the item to be selected into the trained selection model includes: inputting the first feature vector, the second feature vector, and the third feature vector of the item to be selected into the trained selection model.

6. An article selection device, characterized in that, Including: A feature generation module, configured to obtain candidate items, generate the first feature vector of each candidate item according to the click sequence of the candidate item, and generate the second feature vector of each candidate item according to the title of the candidate item; A model training module, configured to determine sample items from the candidate items, and train the selection model according to the first feature vector and the second feature vector of the sample items; A model application module, configured to use the items other than the sample items among the candidate items as the items to be selected, and input the first feature vector and the second feature vector of the items to be selected into the trained selection model; A selection module, configured to determine the target item according to the output item of the trained selection model and the sample item; The feature generation module includes: a category screening unit, configured to: Determine the hot items according to the items clicked in the search results corresponding to the search terms including the hot words within the first preset time period; For each item category, determine the first ratio of the number of hot items included to the total number of categories, and / or determine the second ratio of the sales volume of the hot items included to the total number of categories; According to the first ratio and / or the second ratio, screen out the candidate categories from each item category, and use the items under the candidate categories as candidate items; The model training module includes: a multi-target labeling unit, configured to: Determine the first sample items from the candidate items, where the number of clicks of the first sample items is higher than the first threshold; Determine the second sample items from the candidate items, where the sales volume of the second sample items is higher than the second threshold during the historical hot period; Wherein, the number of the first sample items is less than the number of the second sample items, and the first threshold is greater than the second threshold.

7. A terminal device, characterized in that, The terminal includes: One or more processors; A memory, configured to store 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 item selection method according to any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the item selection method according to any one of claims 1-5.

Images