Product selection method and device based on deep learning

Abstract

This invention discloses a product selection method and apparatus based on deep learning, relating to the field of computer technology. One specific implementation of the method includes: collecting user behavior data, the user behavior data including a list of items corresponding to each search term; constructing item pairs based on the items in the item list to obtain a list of item pairs corresponding to each search term; constructing a product selection model based on a deep learning network according to the list of item pairs corresponding to each search term; and performing product selection based on the product selection model. This implementation, by constructing item pairs, building a product selection model based on the constructed item pairs and the deep learning network, and performing product selection based on the selection model, can greatly expand the model training data and improve the coverage and accuracy of the product selection results.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a product selection method and device based on deep learning. Background Art

[0002] The CPS (Cost Per Sales) system often involves a large number of items. CPS selection is to select and score a large number of CPS items and select the most competitive, high-quality and popular items. Currently, the main methods of CPS selection are: selection based on event registration, data statistics, and manual participation, and selection based on shallow learning. However, the coverage and accuracy of these selection solutions are low. Summary of the invention

[0003] In view of this, an embodiment of the present invention provides a product selection method and device based on deep learning, which can greatly improve the coverage and accuracy of product selection results.

[0004] To achieve the above object, according to one aspect of an embodiment of the present invention, a product selection method based on deep learning is provided, comprising:

[0005] Collecting user behavior data, wherein the user behavior data includes a list of items corresponding to each search term;

[0006] Constructing item pairs according to the items in the item list to obtain an item pair list corresponding to each search term;

[0007] According to the list of item pairs corresponding to each search term, a product selection model is constructed based on a deep learning network, and products are selected according to the product selection model.

[0008] Optionally, constructing item pairs according to the items in the item list to obtain an item pair list corresponding to each search term includes:

[0009] For any search term, the item list corresponding to the search term is taken as the target item list, or the item list after at least part of the items in the item list corresponding to the search term is deleted is taken as the target item list; starting from the first item in the target item list, two adjacent unpaired items are taken as an item pair to obtain the item pair list corresponding to any search term.

[0010] Optionally, deleting at least part of the items in the item list includes: deleting items in the item list whose indicator values ​​of a preset indicator are less than or equal to an indicator threshold.

[0011] Optionally, the method further includes determining a grading label for each item in the item pair list; constructing item pairs based on the items in the item pair list, and after obtaining the item pair list corresponding to each search term, further includes: deleting item pairs in the item pair list that have the same preset grading label.

[0012] Optionally, the user behavior data further includes: the number of clicks and exposures of each item in the item list;

[0013] Determining a grading label for each item in the item pair list includes: determining a click rate for each item in the item pair list according to the click volume and the exposure volume, and determining a grading label for each item in the item pair list according to the click rate.

[0014] Optionally, the click rate of each item in the item pair list is determined according to the click volume and the exposure volume according to the following formula:

[0015]

[0016] In the formula, score represents the click rate of the item, x represents the number of clicks on the item, y represents the number of exposures of the item, and n represents the weight factor.

[0017] Optionally, according to the list of item pairs corresponding to each search term, a product selection model is constructed based on a deep learning network, including:

[0018] All item pairs in the item pair list corresponding to each search term are taken as candidate sample sets, and the candidate sample sets are sampled to obtain training sample sets; according to the training sample sets, a product selection model is constructed based on a deep learning network.

[0019] Optionally, sampling the candidate sample set to obtain a training sample set includes:

[0020] Determine the sampling probability of each item in the candidate sample set, determine the sampling probability of each item pair in the candidate sample set according to the sampling probability of each item, and sample from the candidate sample set according to the sampling probability of each item pair.

[0021] Optionally, the sampling probability of each item in the candidate sample set is determined according to the following formula:

[0022]

[0023] In the formula, P represents the sampling probability of the item, and y represents the exposure of the item.

[0024] Optionally, after sampling the candidate sample set to obtain a training sample set, the method further includes:

[0025] For any item in the candidate sample set, determine whether the indicator value of the preset indicator is abnormal; if so, standardize the indicator value.

[0026] Optionally, determining whether an indicator value of a preset indicator is abnormal includes:

[0027] The index score of the preset index is determined according to the following formula. If the index score is less than or equal to the score threshold, it is determined that there is no abnormality in the index value of the preset index; otherwise, it is determined that there is an abnormality in the index value of the preset index:

[0028]

[0029] In the formula, a′ represents the indicator score of the preset indicator, a represents the indicator value of the preset indicator, mean represents the mean of the preset indicator, and σ represents the standard deviation of the preset indicator.

[0030] Optionally, before building a product selection model based on a deep learning network, the following steps are also included:

[0031] For any item pair in the training sample set, determine whether the click rate of the previous item in the item pair is greater than the click rate of the next item; if so, mark the first sample label for the item pair, otherwise mark the second sample label for the item pair.

[0032] Optionally, the deep learning network adopts a single hidden layer.

[0033] According to a second aspect of an embodiment of the present invention, there is provided a product selection device based on deep learning, comprising:

[0034] A data collection module collects user behavior data, wherein the user behavior data includes a list of items corresponding to each search term;

[0035] An item combination module, which constructs item pairs according to the items in the item list to obtain an item pair list corresponding to each search term;

[0036] The model building module builds a product selection model based on a deep learning network according to a list of item pairs corresponding to each search term, and selects products based on the product selection model.

[0037] Optionally, the item combination module constructs item pairs according to the items in the item list to obtain an item pair list corresponding to each search term, including:

[0038] For any search term, the item list corresponding to the search term is taken as the target item list, or the item list after at least part of the items in the item list corresponding to the search term is deleted is taken as the target item list; starting from the first item in the target item list, two adjacent unpaired items are taken as an item pair to obtain the item pair list corresponding to any search term.

[0039] Optionally, the item combination module is further used to delete at least part of the items in the item list according to the following steps: delete the items in the item list whose index value of a preset index is less than or equal to an index threshold.

[0040] Optionally, the item combination module is further used to: determine a grading label for each item in the item pair list, and after constructing item pairs based on the items in the item list to obtain an item pair list corresponding to each search term, delete the item pairs in the item pair list that have the same preset grading label.

[0041] Optionally, the user behavior data further includes: the number of clicks and exposures of each item in the item list;

[0042] The item combination module determines the grading label of each item in the item pair list, including: determining the click rate of each item in the item pair list according to the click volume and exposure volume, and determining the grading label of each item in the item pair list according to the click rate.

[0043] Optionally, the item grading module determines the click rate of each item in the item pair list according to the click volume and exposure volume according to the following formula:

[0044]

[0045] In the formula, score represents the click rate of the item, x represents the number of clicks on the item, y represents the number of exposures of the item, and n represents the weight factor.

[0046] Optionally, the model building module builds a product selection model based on a deep learning network according to a list of item pairs corresponding to each search term, including:

[0047] All item pairs in the item pair list corresponding to each search term are taken as candidate sample sets, and the candidate sample sets are sampled to obtain training sample sets; according to the training sample sets, a product selection model is constructed based on a deep learning network.

[0048] Optionally, the model building module samples the candidate sample set to obtain a training sample set, including:

[0049] Determine the sampling probability of each item in the candidate sample set, determine the sampling probability of each item pair in the candidate sample set according to the sampling probability of each item, and sample from the candidate sample set according to the sampling probability of each item pair.

[0050] Optionally, the model building module determines the sampling probability of each item in the candidate sample set according to the following formula:

[0051]

[0052] In the formula, P represents the sampling probability of the item, and y represents the exposure of the item.

[0053] Optionally, after the model building module samples the candidate sample set to obtain a training sample set, it further includes:

[0054] For any item in the candidate sample set, determine whether the indicator value of the preset indicator is abnormal; if so, standardize the indicator value.

[0055] Optionally, the model building module determines whether there is an abnormality in the indicator value of the preset indicator, including:

[0056] The index score of the preset index is determined according to the following formula. If the index score is less than or equal to the score threshold, it is determined that there is no abnormality in the index value of the preset index; otherwise, it is determined that there is an abnormality in the index value of the preset index:

[0057]

[0058] In the formula, a′ represents the indicator score of the preset indicator, a represents the indicator value of the preset indicator, mean represents the mean of the preset indicator, and σ represents the standard deviation of the preset indicator.

[0059] Optionally, before the model building module builds the product selection model based on the deep learning network, it is also used to:

[0060] For any item pair in the training sample set, determine whether the click rate of the previous item in the item pair is greater than the click rate of the next item; if so, mark the first sample label for the item pair, otherwise mark the second sample label for the item pair.

[0061] Optionally, the deep learning network adopts a single hidden layer.

[0062] According to a third aspect of an embodiment of the present invention, there is provided an electronic device for selecting products based on deep learning, including:

[0063] one or more processors;

[0064] a storage device for storing one or more programs,

[0065] When the one or more programs are executed by the one or more processors, the one or more processors implement the method provided in the first aspect of the embodiment of the present invention.

[0066] According to a fourth aspect of an embodiment of the present invention, a computer-readable medium is provided, on which a computer program is stored, and when the program is executed by a processor, the method provided by the first aspect of the embodiment of the present invention is implemented.

[0067] One embodiment of the above invention has the following advantages or beneficial effects: by constructing item pairs, constructing a product selection model based on the constructed item pairs and a deep learning network, and selecting products according to the product selection model, the model training data can be greatly expanded and the coverage and accuracy of the product selection results can be improved.

[0068] The further effects of the above-mentioned non-conventional optional manner will be described below in conjunction with the specific implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The accompanying drawings are used to better understand the present invention and do not constitute an improper limitation of the present invention.

[0070] Figure 1 It is a schematic diagram of the main process of the product selection method based on deep learning in an embodiment of the present invention;

[0071] Figure 2 It is a flowchart of a product selection method based on deep learning in an optional embodiment of the present invention;

[0072] Figure 3 is a schematic diagram of main modules of a product selection device based on deep learning according to an embodiment of the present invention;

[0073] Figure 4 is an exemplary system architecture diagram to which embodiments of the present invention may be applied;

[0074] Figure 5 It is a schematic diagram of the structure of a computer system of a terminal device or a server suitable for implementing an embodiment of the present invention. DETAILED DESCRIPTION

[0075] The following is a description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0076] According to one aspect of an embodiment of the present invention, a product selection method based on deep learning is provided.

[0077] Figure 1 is a schematic diagram of the main process of the product selection method based on deep learning in an embodiment of the present invention, such as Figure 1 As shown, the product selection method based on deep learning includes: step S101, step S102, and step S103.

[0078] Step S101: Collect user behavior data.

[0079] User behavior data is data obtained by detecting user behavior. User behavior data may include data on behaviors such as searching, browsing, clicking, purchasing, and placing orders. Optionally, user behavior data includes a list of items corresponding to each search term. Of course, user behavior data may also include data on various indicators for evaluating user behavior, such as the exposure of each item in the item list, the number of clicks by users on each item in the item list, etc.

[0080] The time span of user behavior data can be selectively set according to actual conditions. For example, collecting user behavior data within a larger time span, such as all user behavior data; or collecting user behavior data within a smaller time span, such as user behavior data within only one month.

[0081] In actual application, a time span threshold can be set so that the time span of the collected user behavior data is not less than the time span threshold, such as 7 days, 30 days, etc. For example, in some scenarios, user behavior data is relatively sensitive to time. For example, the difference between user behavior data on weekends and weekdays is relatively large. The time span threshold can be set to 7 days. When collecting data, user behavior data of at least the most recent 7 days is collected. By setting the time span threshold, the accuracy of product selection results can be improved.

[0082] Step S102: construct item pairs according to the items in the item list to obtain an item pair list corresponding to each search term.

[0083] The present invention constructs item pairs based on the Pairwise algorithm, and finally constructs a product selection model through a deep learning method to select items. This method can not only greatly expand the training data, but also clearly identify long-tail items that cannot be covered by shallow learning, and at the same time, it can use the nonlinear characteristics and high-precision characteristics of deep learning to meet the needs of complex business scenarios.

[0084] Optionally, item pairs are constructed based on the items in the item list to obtain a list of item pairs corresponding to each search term, including: for any search term, taking the item list corresponding to the search term as the target item list, or taking the list after deleting at least part of the items in the item list corresponding to the search term as the target item list; starting from the first item in the target item list, taking two adjacent unpaired items as an item pair to obtain a list of item pairs corresponding to any search term.

[0085] By treating two adjacent unpaired items as an item pair for deep learning, we can fully learn the significant and non-significant differences between items, especially long-tail items with relatively low exposure, and improve the coverage and accuracy of product selection results.

[0086] Optionally, deleting at least part of the items in the item list includes: deleting the items in the item list whose index value of the preset index is less than or equal to the index threshold. The preset index refers to a pre-set index, which can be selectively set according to actual conditions, such as exposure, click-through rate, sales volume, etc. The index threshold can be selectively set according to actual conditions, such as being set to a fixed value, or determined according to the index value of the preset index of each item.

[0087] In an optional embodiment, the quartile of the index value of the preset index of all items in the item list corresponding to each search term is used as the index threshold. When determining the quartile, the index values ​​of the preset index of all items in the item list corresponding to each search term can be first arranged in descending order and divided into four equal parts, and then the index value of the item at the first division point is taken as the quartile. In the field of e-commerce, when exposure is used as the preset index, the quartile is generally around 6.

[0088] Deleting items in the item list whose index values ​​of the preset index are less than or equal to the index threshold can reduce the amount of calculation while improving the accuracy of the item selection results.

[0089] Optionally, the method further includes determining a grading label for each item in the item pair list; constructing item pairs based on the items in the item pair list, and after obtaining the item pair list corresponding to each search term, further includes: deleting item pairs in the item pair list that have the same preset grading label.

[0090] Grading labels are used to distinguish different grades of items, and the names and numbers of grading labels can be selectively set according to actual conditions. For example, three grading labels of "superior", "medium", and "poor" are set, or four grading labels of "first class", "second class", "third class", and "fourth class" are set. The preset grading label refers to a grading label among multiple grading labels. Exemplarily, when the three grading labels of "superior", "medium", and "poor" are set, after obtaining the item pair lists corresponding to each search term, the grading labels of the two items in the same item pair may be "superior" and "superior", "medium" and "medium", "poor" and "poor", "superior" and "poor", "medium" and "poor", "superior" and "medium". Assuming that the "poor" grading label is set as the preset grading label, after obtaining the item pair lists corresponding to each search term, the item pairs with the grading labels of "poor" and "poor" in the item pair list are deleted.

[0091] In this example, the step of determining the grading label of each item in the item pair list can be performed before the step of constructing the item pair, or simultaneously with the step of determining the item pair list corresponding to the search term, or after obtaining the item pair list corresponding to each search term. Deleting item pairs with the same preset grading label in the item pair list can reduce the amount of calculation and improve the accuracy of the product selection results.

[0092] The grading label of each item can be determined manually based on experience, or based on the indicator value of one or more indicators. Optionally, the user behavior data also includes: the number of clicks and exposures of each item in the item list. Determining the grading label of each item in the item pair list includes: determining the click rate of each item in the item pair list based on the click rate and exposure, and determining the grading label of each item in the item pair list based on the click rate.

[0093] Exposure refers to the number of times an item is seen by users. No matter on which page or in which scene, as long as the item is seen by users once, the exposure of the item will increase once. Clicks refer to the number of times an item is clicked by users. Generally speaking, the higher the click-through rate, the more times the item is clicked under the same exposure. Therefore, the click-through rate can reflect user preferences to a certain extent. Determining the grading label of an item based on the click-through rate can make the product selection results more in line with user preferences and improve the accuracy of the product selection results.

[0094] When determining the click-through rate, the click-through rate of the item may be obtained by dividing the number of clicks on the item by the number of exposures of the item. Optionally, the click-through rate of each item in the item pair list may be determined according to the number of clicks and the number of exposures, according to the following formula 1:

[0095]

[0096] In the formula, score represents the click rate of the item, x represents the number of clicks on the item, y represents the number of exposures of the item, and n represents the weight factor. The value of the weight factor can be selectively set according to the actual situation. Usually, the weight factor can be set to 1.1. Those skilled in the art can also set it to other values ​​according to the actual situation, such as 1.2, 0.8, etc. The weight factors of different items can be the same or different. The summation of formula 1 refers to the summation of the corresponding values ​​of the two items in the item pair.

[0097] Compared with the method of obtaining the click-through rate of an item by dividing the number of clicks on the item by the number of exposures of the item, the click-through rate determined by the above formula is closer to the optimal solution for sorting application technologies such as search. The subsequent steps of constructing a product selection model based on the click-through rate determined by the above formula can improve the accuracy of the product selection results.

[0098] Step S103: According to the list of item pairs corresponding to each search term, a product selection model is constructed based on a deep learning network, and products are selected according to the product selection model.

[0099] In this step, features can be extracted and constructed based on the static attributes of the item (such as price, commission, etc.), activity information (such as group buying, flash sales, etc.), store data (such as sales, DSR (Detail Seller Rating, seller service rating system), store level, etc.), statistical attributes of the item (such as sales, commission changes, etc.), classification statistical data (such as popularity, sales ranking, etc.), coupon attributes and other related attributes. Then, the product selection method is constructed through deep learning methods. Deep learning combines low-level features to form more abstract high-level representation attribute categories or features to discover distributed feature representations of data. The specific method of deep learning can be selected according to the actual situation, such as DBN (Deep belief network), CNN (Convolution Neural Networks), RNN (Recurrent neural network), etc.

[0100] Through the nonlinear characteristics of deep learning, the parameters of each neural network are continuously trained to build a product selection model. Finally, the product selection model obtained through deep learning predicts the probability of each item being clicked, and then product selection is carried out.

[0101] When building a product selection model based on a deep learning network, batch training can be used for training. For example, all item pairs under the same search term are substituted into the neural network for forward feedback, and then the total difference (that is, the difference between the actual probability of being clicked on the item pair and the predicted value) is calculated and the error is back-propagated. This will greatly reduce the number of error back-propagations.

[0102] The neural network of deep learning is divided into three layers: input layer, hidden layer and output layer. The loss function can be:

[0103]

[0104]

[0105] In the formula, S i S represents the true probability of the item pair consisting of item i and item j being clicked; i represents the true probability of item i being clicked, S j represents the actual probability of item j being clicked, and its calculation method is shown in Formula 5; S i >S j Time ij =1, otherwise S ij =0; f() represents the activation function, the content in the brackets represents the variable of the activation function, a represents the input of each layer, that is, the engineering information input related to the item, θ represents the activation threshold; K represents the number of input layers, M represents the number of hidden units in the hidden layer; f m () represents the activation function of the mth hidden unit. The content in brackets represents the variable of the activation function of the mth hidden unit. k represents the input of the kth input layer, w mk represents the weight from the input layer to the hidden layer of the mth hidden unit of the kth input layer, b m represents the bias from the input layer to the hidden layer of the mth hidden unit in the kth input layer, w m represents the weight from the hidden layer to the output layer of the mth hidden unit of the kth input layer, and b represents the bias from the hidden layer to the output layer.

[0106] Optionally, according to the item pair list corresponding to each search term, a product selection model is constructed based on a deep learning network, including: taking all item pairs in the item pair list corresponding to each search term as a candidate sample set, sampling the candidate sample set to obtain a training sample set; and constructing a product selection model based on the deep learning network according to the training sample set. Sampling the candidate sample set to obtain a training sample set can increase the randomness of the samples and the probability of selecting items with high index values ​​(such as exposure), thereby improving the accuracy of the product selection results.

[0107] Optionally, sampling the candidate sample set to obtain a training sample set includes: determining the sampling probability of each item in the candidate sample set, determining the sampling probability of each item pair in the candidate sample set according to the sampling probability of each item, and sampling from the candidate sample set according to the sampling probability of each item pair. In this example, the sampling probability of an item pair is the product of the sampling probabilities of the two items in the item pair. By separately determining the sampling probability of each item pair, the accuracy of the product selection result can be further improved.

[0108] Optionally, the sampling probability of each item in the candidate sample set is determined according to the following formula 2:

[0109]

[0110] In the formula, P represents the sampling probability of the item, and y represents the exposure of the item. Using the formula in this example to determine the sampling probability of the item can significantly improve the accuracy of the product selection results.

[0111] Optionally, after sampling the candidate sample set to obtain the training sample set, the method further includes: for any item in the candidate sample set, determining whether the index value of the preset index is abnormal; if so, standardizing the index value. Standardization refers to replacing the abnormal value with a non-abnormal value, such as replacing it with the average value, median, etc. of the preset index. Standardizing the abnormal index value can avoid the influence of the abnormal value on the product selection result and improve the accuracy of the product selection result.

[0112] The basis for determining abnormal values ​​can be selectively set according to actual conditions. For example, if the number of clicks on a certain item suddenly increases while other conditions remain unchanged, the item is likely to have fake orders, and the number of clicks on the item can be determined as an abnormal value. Optionally, determining whether the indicator value of a preset indicator is abnormal includes:

[0113] The index score of the preset index is determined according to the following formula 3. If the index score is less than or equal to the score threshold, it is determined that there is no abnormality in the index value of the preset index; otherwise, it is determined that there is an abnormality in the index value of the preset index:

[0114]

[0115] In the formula, a′ represents the index score of the preset index, a represents the index value of the preset index, mean represents the average of the preset index, and σ represents the standard deviation of the preset index. The score threshold can be selectively set according to the actual situation, for example, to 3.5, 5, etc. The above formula can greatly improve the recognition effect of outliers and improve the accuracy of the final product selection results.

[0116] Optionally, before building a product selection model based on a deep learning network, the method further includes: for any item pair in the training sample set, determining whether the click rate of the first item in the item pair is greater than the click rate of the second item; if so, marking a first sample label for the item pair, otherwise marking a second sample label for the item pair. Exemplarily, the first sample label is 1, representing a positive sample, and the second sample label is 0, representing a negative sample; or, the first sample label is 0, representing a negative sample, and the second sample label is 1, representing a positive sample.

[0117] The present invention obtains training samples based on item pairs, and the distribution of sample data is relatively wide as a whole, so a neural network with a single hidden layer can meet the relevant method training requirements. Therefore, optionally, the deep learning network adopts a single hidden layer. Using a single hidden layer can reduce the amount of calculation while meeting the accuracy. In other optional embodiments, the number of hidden units in the hidden layer is 15. If the number of hidden units is higher or lower than 15, overfitting or low accuracy will occur. Using 15 hidden units can avoid overfitting while meeting the accuracy.

[0118] The following combination Figure 2 The product selection method of the embodiment of the present invention is exemplified. The scenario of this embodiment is to collect user behavior data in related applications of CPS advertising, and finally construct a product selection method from the user behavior data. The product selection scenario is relatively complex, and various factors need to be considered. In addition to the item's own factors such as commission, price, label, style, color, sales volume, classification, etc., it is also related to various non-item factors such as the promotion activities that the item has recently registered for, the comprehensive score of the merchant or store, brand awareness, current popular elements, etc. Of course, in addition to the algorithm, it is also necessary to consider factors of actual commercial applications, such as e-commerce order brushing, banned investment categories, etc. These factors will also interfere with conventional product selection algorithms, especially some temporary business policies.

[0119] The general practice is to collect the click-through rate (CTR) data of the item to determine the positive and negative samples, and then construct the machine learning training set algorithm for the positive and negative samples according to the level of the CTR data. Because of the long tail phenomenon (also known as the long tail effect, that is, most of the demand will be concentrated in the head), the exposure and click volume of most CPS items are very low. Statistically speaking, the CTR calculated because of the low traffic is not the real CTR of the item, so these items belong to items with fuzzy positive and negative boundaries. This approach easily causes most of the items with fuzzy positive and negative boundaries to be unable to participate in the training. At the same time, this method is limited by the number of CTR data (because items with obvious positive and negative sample distinctions are often concentrated in a few item categories, and simply expanding the sampled items cannot enrich the positive and negative samples). Often, due to the reduction of training samples, it is impossible to truly train an accurate model or an overly simple model, resulting in the inability to truly meet the complex personalized advertising needs of CPS. This embodiment proposes a method based on the Pairwise algorithm (a method of constructing item pairs by CTR size) to construct positive and negative samples, and finally constructs an algorithm through a deep learning method to select products. This method can not only greatly expand the training data, but also clearly identify long-tail items that shallow learning cannot cover, and at the same time use the nonlinear characteristics and high-precision characteristics of deep learning to meet the needs of complex business scenarios. Figure 2 , the specific method is as follows:

[0120] Step S201: Collect user behavior data for at least 7 days, wherein the user behavior data includes: a list of items corresponding to each search term;

[0121] Step S202: Determine the click rate of each item in the item pair list according to the number of clicks and the number of exposures, and refer to Formula 1 for the calculation formula above;

[0122] Step S203, determining a grading label for each item in the item pair list according to the click rate; marking items with a CTR≤0.023 with a grading label of "poor", marking items with a CTR>0.023 and CTR<0.06 with a grading label of "medium", and marking items with a CTR≥0.06 with a grading label of "superior";

[0123] Step S204: construct item pairs according to the items in the item list to obtain a candidate sample set; the construction method is described in the above related text and will not be repeated here;

[0124] Step S205: sampling the candidate sample set to obtain a training sample set; the sampling method is described in the above related text and will not be repeated here;

[0125] Step S206: normalize the abnormal index values ​​of the preset indexes; the normalization method is described in the above text and will not be described again here;

[0126] Step S207: construct a product selection model based on the deep learning network; the construction method is described in the previous text and will not be repeated here.

[0127] The present invention can eliminate the interference of actual scenarios on the algorithm to simultaneously improve the coverage and accuracy of the algorithm for CPS selection.

[0128] The main problems with traditional methods based on event registration, data statistics, and manual participation include:

[0129] 1) Compared with the huge set of candidate items, only a few items can be selected through active items and manual inspection;

[0130] 2) Artificial experience is not necessarily correct (the characteristics of objects will change with time, region, and scene, which are areas that artificial experience cannot reach);

[0131] 3) Most long-tail items with hot-selling potential cannot be identified;

[0132] The problems with the method of selecting products based on shallow learning include:

[0133] 1) It cannot cover many practical comprehensive application factors such as personalized advertising and commercial activities, multi-dimensional item attributes, complex application scenarios, and personalized media;

[0134] 2) The distribution of extracted features in the feature space is relatively sparse. For example, there are often hundreds of millions of CPS items, but the number of items sold in a single day is often less than one million. Therefore, the sales volume of most items in the sales volume feature is 0, which means that the data is sparsely distributed.

[0135] 3) The performance of the classifier is greatly affected by the static properties of the training items (such as price changes, commission increases, etc.);

[0136] 4) The positive and negative samples during offline training cannot cover most real application scenarios, and most samples are distributed at the fuzzy boundary between positive and negative samples;

[0137] 5) Some items have unique application scenario personalization and special knowledge attributes, and conventional data mining algorithms cannot identify these features;

[0138] 6) The particularity of the CPS ecosystem, such as the deviation between CPS users’ preferences for commissions, coupons, new products, etc. and the benefits brought by actual items, will interfere with the model.

[0139] The problems of the above two methods lead to low coverage and accuracy in product selection. The present invention can greatly expand the model training data and improve the coverage and accuracy of product selection results by constructing item pairs, constructing a product selection model based on the constructed item pairs and a deep learning network, and selecting products according to the product selection model.

[0140] The product selection method of the embodiment of the present invention can be used for search and recommendation related default item recall, and can also be combined with search, recommendation, advertising push and other technologies, and applied to CPS related advertising, search recommendation, social e-commerce, media cooperation and other fields.

[0141] According to a second aspect of an embodiment of the present invention, a device for implementing the above method is provided.

[0142] Figure 3 is a schematic diagram of the main modules of the deep learning-based product selection device according to an embodiment of the present invention. Figure 3 As shown, the product selection device 300 based on deep learning includes:

[0143] A data collection module 301 collects user behavior data, wherein the user behavior data includes a list of items corresponding to each search term;

[0144] An item combination module 302 constructs item pairs according to the items in the item list to obtain an item pair list corresponding to each search term;

[0145] The model building module 303 builds a product selection model based on a deep learning network according to a list of item pairs corresponding to each search term, and selects products according to the product selection model.

[0146] Optionally, the item combination module constructs item pairs according to the items in the item list to obtain an item pair list corresponding to each search term, including:

[0147] For any search term, the item list corresponding to the search term is taken as the target item list, or the item list after at least part of the items in the item list corresponding to the search term is deleted is taken as the target item list; starting from the first item in the target item list, two adjacent unpaired items are taken as an item pair to obtain the item pair list corresponding to any search term.

[0148] Optionally, the item combination module is further used to delete at least part of the items in the item list according to the following steps: delete the items in the item list whose index value of a preset index is less than or equal to an index threshold.

[0149] Optionally, the item combination module is further used to: determine a grading label for each item in the item pair list, and after constructing item pairs based on the items in the item list to obtain an item pair list corresponding to each search term, delete the item pairs in the item pair list that have the same preset grading label.

[0150] Optionally, the user behavior data further includes: the number of clicks and exposures of each item in the item list;

[0151] The item combination module determines the grading label of each item in the item pair list, including: determining the click rate of each item in the item pair list according to the click volume and exposure volume, and determining the grading label of each item in the item pair list according to the click rate.

[0152] Optionally, the item grading module determines the click rate of each item in the item pair list according to the click volume and exposure volume according to the following formula:

[0153]

[0154] In the formula, score represents the click rate of the item, x represents the number of clicks on the item, y represents the number of exposures of the item, and n represents the weight factor.

[0155] Optionally, the model building module builds a product selection model based on a deep learning network according to a list of item pairs corresponding to each search term, including:

[0156] All item pairs in the item pair list corresponding to each search term are taken as candidate sample sets, and the candidate sample sets are sampled to obtain training sample sets; according to the training sample sets, a product selection model is constructed based on a deep learning network.

[0157] Optionally, the model building module samples the candidate sample set to obtain a training sample set, including:

[0158] Determine the sampling probability of each item in the candidate sample set, determine the sampling probability of each item pair in the candidate sample set according to the sampling probability of each item, and sample from the candidate sample set according to the sampling probability of each item pair.

[0159] Optionally, the model building module determines the sampling probability of each item in the candidate sample set according to the following formula:

[0160]

[0161] In the formula, P represents the sampling probability of the item, and y represents the exposure of the item.

[0162] Optionally, after the model building module samples the candidate sample set to obtain a training sample set, it further includes:

[0163] For any item in the candidate sample set, determine whether the indicator value of the preset indicator is abnormal; if so, standardize the indicator value.

[0164] Optionally, the model building module determines whether there is an abnormality in the indicator value of the preset indicator, including:

[0165] The index score of the preset index is determined according to the following formula. If the index score is less than or equal to the score threshold, it is determined that there is no abnormality in the index value of the preset index; otherwise, it is determined that there is an abnormality in the index value of the preset index:

[0166]

[0167] In the formula, a′ represents the indicator score of the preset indicator, a represents the indicator value of the preset indicator, mean represents the mean of the preset indicator, and σ represents the standard deviation of the preset indicator.

[0168] Optionally, before the model building module builds the product selection model based on the deep learning network, it is also used to:

[0169] For any item pair in the training sample set, determine whether the click rate of the previous item in the item pair is greater than the click rate of the next item; if so, mark the first sample label for the item pair, otherwise mark the second sample label for the item pair.

[0170] Optionally, the deep learning network uses a single hidden layer, and the number of hidden units in the hidden layer is 15

[0171] According to a third aspect of an embodiment of the present invention, there is provided an electronic device for selecting products based on deep learning, including:

[0172] one or more processors;

[0173] a storage device for storing one or more programs,

[0174] When the one or more programs are executed by the one or more processors, the one or more processors implement the method provided in the first aspect of the embodiment of the present invention.

[0175] According to a fourth aspect of an embodiment of the present invention, a computer-readable medium is provided, on which a computer program is stored, and when the program is executed by a processor, the method provided by the first aspect of the embodiment of the present invention is implemented.

[0176] Figure 4An exemplary system architecture 400 is shown to which a deep learning-based product selection method or a deep learning-based product selection device according to an embodiment of the present invention can be applied.

[0177] like Figure 4 As shown, system architecture 400 may include terminal devices 401, 402, 403, network 404 and server 405. Network 404 is used to provide a medium for communication links between terminal devices 401, 402, 403 and server 405. Network 404 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0178] Users can use terminal devices 401, 402, 403 to interact with server 405 through network 404 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 401, 402, 403, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (only as examples).

[0179] The terminal devices 401 , 402 , and 403 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, and desktop computers.

[0180] The server 405 may be a server that provides various services, such as a backend management server (only an example) that provides support for shopping websites browsed by users using the terminal devices 401, 402, and 403. The backend management server may analyze and process the received data such as product information query requests, and feed back the processing results (such as target push information, product information - only an example) to the terminal device.

[0181] It should be noted that the deep learning-based product selection method provided in the embodiment of the present invention is generally executed by the server 405 , and accordingly, the deep learning-based product selection device is generally arranged in the server 405 .

[0182] It should be understood that Figure 4 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0183] Reference below Figure 5 , which shows a schematic diagram of the structure of a computer system 500 of a terminal device suitable for implementing an embodiment of the present invention. Figure 5 The terminal device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0184] like Figure 5As shown, the computer system 500 includes a central processing unit (CPU) 501, which can 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 part 508 into a random access memory (RAM) 503. In the RAM 503, various programs and data required for the operation of the system 500 are also stored. The CPU 501, the ROM 502, and the RAM 503 are connected to each other via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0185] The following components are connected to the I / O interface 505: an input section 506 including a keyboard, a mouse, etc.; an output section 507 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN card, a modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to the I / O interface 505 as needed. A removable medium 511, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 510 as needed, so that a computer program read therefrom is installed into the storage section 508 as needed.

[0186] In particular, according to the embodiments disclosed in the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 509, and / or installed from the removable medium 511. When the computer program is executed by the central processing unit (CPU) 501, the above-mentioned functions defined in the system of the present invention are executed.

[0187] It should be noted that the computer-readable medium shown in the present invention may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present invention, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0188] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0189] The modules involved in the embodiments of the present invention may be implemented in software or hardware. The modules described may also be set in a processor, for example, may be described as: a processor comprising: a data acquisition module, which collects user behavior data, wherein the user behavior data comprises: a list of items corresponding to each search term; an item combination module, which constructs item pairs based on the items in the item list, and obtains a list of item pairs corresponding to each search term; a model construction module, which constructs a product selection model based on a deep learning network according to the list of item pairs corresponding to each search term, and selects products based on the product selection model. Among them, the names of these modules do not constitute a limitation on the modules themselves in some cases. For example, the data acquisition module may also be described as a "module for constructing a product selection model based on a deep learning network."

[0190] As another aspect, the present invention further provides a computer-readable medium, which may be included in the device described in the above embodiment; or it may exist independently without being assembled into the device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by a device, the device includes: collecting user behavior data, the user behavior data includes: a list of items corresponding to each search term; constructing item pairs based on the items in the item list to obtain a list of item pairs corresponding to each search term; constructing a product selection model based on a deep learning network according to the list of item pairs corresponding to each search term, and selecting products based on the product selection model.

[0191] According to the technical solution of the embodiments of the present invention, by constructing item pairs, constructing a product selection model based on the constructed item pairs and a deep learning network, and performing product selection according to the product selection model, the model training data can be greatly expanded and the coverage and accuracy of the product selection results can be improved.

[0192] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A product selection method based on deep learning, characterized in that: include: Collecting user behavior data, wherein the user behavior data includes a list of items corresponding to each search term; Constructing item pairs according to the items in the item list to obtain an item pair list corresponding to each search term; According to the list of item pairs corresponding to each search term, a product selection model is constructed based on a deep learning network, and products are selected according to the product selection model; Item pairs are constructed based on the items in the item list to obtain an item pair list corresponding to each search term, including: for any search term, taking the item list corresponding to the search term as the target item list, or taking the list after deleting at least a part of the items in the item list corresponding to the search term as the target item list; starting from the first item in the target item list, taking two adjacent unpaired items as an item pair to obtain an item pair list corresponding to any search term.

2. The method according to claim 1, characterized in that Deleting at least a portion of the items in the item list includes: deleting the items in the item list whose index values ​​of a preset index are less than or equal to an index threshold.

3. The method according to claim 1, characterized in that Also includes: determining a grading label for each item in the item pair list; After constructing item pairs according to the items in the item list and obtaining the item pair lists corresponding to the search terms, the method further includes: deleting item pairs in the item pair list that have the same preset grading label.

4. The method according to claim 3, characterized in that The user behavior data also includes: the number of clicks and exposures of each item in the item list; Determining a grading label for each item in the item pair list includes: determining a click rate for each item in the item pair list according to the click volume and the exposure volume, and determining a grading label for each item in the item pair list according to the click rate.

5. The method according to claim 4, characterized in that According to the list of item pairs corresponding to each search term, a product selection model is built based on a deep learning network, including: All item pairs in the item pair list corresponding to each search term are taken as candidate sample sets, and the candidate sample sets are sampled to obtain training sample sets; according to the training sample sets, a product selection model is constructed based on a deep learning network.

6. The method according to claim 5, characterized in that Sampling the candidate sample set to obtain a training sample set includes: Determine the sampling probability of each item in the candidate sample set, determine the sampling probability of each item pair in the candidate sample set according to the sampling probability of each item, and sample from the candidate sample set according to the sampling probability of each item pair.

7. The method according to claim 5, characterized in that After sampling the candidate sample set to obtain a training sample set, the method further includes: For any item in the candidate sample set, determine whether the indicator value of the preset indicator is abnormal; if so, standardize the indicator value.

8. The method according to claim 5, characterized in that Before building a product selection model based on a deep learning network, it also includes: For any item pair in the training sample set, determine whether the click rate of the previous item in the item pair is greater than the click rate of the next item; if so, mark the first sample label for the item pair, otherwise mark the second sample label for the item pair.

9. The method according to claim 1, characterized in that The deep learning network adopts a single hidden layer.

10. A product selection device based on deep learning, characterized in that: include: A data collection module, used to collect user behavior data, wherein the user behavior data includes a list of items corresponding to each search term; An item combination module, which constructs item pairs according to the items in the item list to obtain an item pair list corresponding to each search term; A model building module, which builds a product selection model based on a deep learning network according to a list of item pairs corresponding to each search term, and selects products according to the product selection model; Item pairs are constructed based on the items in the item list to obtain an item pair list corresponding to each search term, including: for any search term, taking the item list corresponding to the search term as the target item list, or taking the list after deleting at least a part of the items in the item list corresponding to the search term as the target item list; starting from the first item in the target item list, taking two adjacent unpaired items as an item pair to obtain an item pair list corresponding to any search term.

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

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

Images