JP2012529122A - System and method for learning user genre and style and matching products to user preferences - Google Patents

Abstract

  A user's fashion preferences are determined based on user interactions with a plurality of fashion product content items that individually indicate the corresponding fashion product. Make recommendations on the fashion product to the user based at least in part on the user's fashion preferences.

Description

  The present invention relates to a system and method for learning a user's genre and style and matching a product to the user's preferences.

  Digital photography has become a very important consumer application. As a result, we enjoy the convenience of digital image capture and sharing. The price of devices that capture digital images is declining, and the ability to send photos from one location to another is one of the driving forces for network bandwidth expansion.

  Due to the relatively low cost of memory and the availability of devices and platforms that can view digital images, the average consumer is hard drives, CD-ROMs, and flash Most of the digital images are stored on a computer readable medium such as a memory. For organizing, file folders are mainly used, but applications that assist users in organizing and browsing digital images have also been created.

  Online learning is a machine learning paradigm where one algorithm learns from one instance or sample at a time. While off-line learning consists of well-established and well-established techniques, over the past decade, online algorithms have received a great deal of attention, learning complex background and outline models, object detection and There are several diverse applications such as classification, user behavior modeling and prediction. Online learning is the only possible solution in applications where training data is absolutely not available in batch form, but is collected at the same time as the decision / classification process, and therefore needs to design an adaptive learning method Can be. In contrast, in the offline or batch paradigm, retraining is required when new / unknown data is presented.

  In the literature, several on-line variants of the most common off-line machine learning algorithms have been proposed. Some methods attempt to address the training problem of Support Vector Machine (SVM) with large amounts of data. SVM training requires the quadratic programming to be solved at a number of factors equal to the concentration in the training set, which can impede memory requirements and therefore requires an on-line alternative . One method attempts to introduce an incremental decision tree classifier that can be updated and retrained using new unknown data instances. Proposals to extend the general AdaBoost algorithm to an online scenario have been made in several literatures, and there are several interesting variations ranging from semi-supervised boosting to multi-instance learning.

  In fashion, people usually have their own unique preferences for clothing styles, colors, and genres. People's preferences regarding genres and styles are nurtured by their personal experiences. For example, in an actual store, the customer can explain to the salesperson about their preferences and styles, so that the salesperson can use the appropriate set of clothing and fashion accessories that match the customer's preferences. Can be recommended to customers. However, embodiments recognize that this is not the case for online shopping. In online shopping, customers must search and scan a large number of products for matching styles and preferences.

  Accordingly, the embodiments described herein provide a computer-implemented method or system that allows a program to determine a user's genre preference for style or fashion.

  Furthermore, according to the embodiment, classification and categorization by a program of fashion products using images, texts, and metadata associated with the corresponding fashion product content items are possible.

  Further, according to some embodiments, the service or system may be in a determined genre of the style of the fashion product represented by the information and / or content item determined in relation to the user's genre preference. Based on this, the recommendations determined by the program for the top fashion product can be implemented.

  More particularly, the embodiments described herein include computer-implemented methods for determining user preferences for fashion products. In one embodiment, the user's fashion preferences are determined based on user interactions with a plurality of fashion product content items that individually indicate the corresponding fashion product. Make recommendations for fashion products to the user based at least in part on the user's fashion preferences.

  According to another embodiment, the fashion product content item is analyzed to determine a set of features of the fashion product indicated in the fashion product content item. Based on the quantitative analysis of the determined set of features, the fashion product is associated with a predefined description category for each of the plurality of description categories. Product content items and their predefined description categories for each of a plurality of description categories are used to determine or predict user preferences.

  In another embodiment, one or more processors (such as those provided in any computing environment such as a server-client) analyze individual fashion product content items that represent a catalog of fashion products, and each Each fashion product content item is configured to determine a set of features of the fashion product shown in the fashion product content item. Each fashion product represented by one of the fashion product content items is assigned to one or more corresponding description categories of predefined description categories. This assignment is based on a quantitative analysis of the determined set of features. Detect one or more fashion product content items that may be of interest to the user. A predefined descriptive category for each of a plurality of descriptive categories of one or more fashion product content items that may be of interest to the user is used to determine the user's fashion preferences.

  Embodiments described herein include (i) learning user or customer preferences in clothing style, fashion, genre, and (ii) predicting genres of different clothing products and fashion accessories. And / or (iii) (a) the user's known shopping parameters (eg, user size information, price preferences, likes and dislikes for particular styles, patterns, and colors) and / or (b) each individual user's Includes systems and methods that use predicted genres and styles to suggest the best matching products and accessories to customers.

  Fashion products include, for example, clothing items, accessories, and apparel. Specific examples include blouses, shirts, dresses, shoes, socks, pants and bottoms, belts, jewelry (eg, watches, earrings, necklaces), ties, hats, jackets, and coats.

  A fashion product content item corresponds to a document or file that contains visual, textual, and / or metadata information about a particular product. Fashion product content items are typically provided as part of an online catalog or e-commerce search engine. Exemplary aspects of such content items include (i) one or more images of the product, (ii) text information about the product including information about available sizes and variations of the product, (iii) price information, and And / or (iv) include links or data elements that allow a viewer of the content item to easily purchase the presented fashion product.

  In all of these online learning methods, computational complexity and latency issues remain unresolved and time limited, such as real-time object tracking or the online shopping scenarios described herein. Some embodiments recognize that it can be very important in certain applications. In fact, in order to meet strict time requirements, strict requirements are imposed on the number of operations allowed by a large number of higher-order feature vectors.

  Some embodiments described herein are computer-implemented for learning user preferences from user interactions with an online interface (eg, such as that provided on a shopping website) Including the law. By predicting what the user likes, a better search ranking algorithm can be designed, which results in the user having a better experience. In terms of feature selection, embodiments combine disparate cues derived from visual features and text features, and in particular features that are compact to the user's preference that traditional features cannot be realized. Provide a realistic expression. Furthermore, embodiments implement a learning phase that can process relatively large feature vectors in less than a few milliseconds to avoid compromising the overall user experience.

  As used herein, the terms “by program” and “by program”, or variations thereof, mean by implementation of code, programming, or other logic. Programmatic operations may be performed by software, firmware, or hardware, and generally without user intervention. Since the operation may be triggered manually, it is not essential but may be executed automatically.

  One or more embodiments described herein may be implemented using program elements, which are often referred to as modules or components, although other names may be used. May be. Such program elements may include programs, subroutines, portions of programs, or software or hardware components capable of performing one or more described tasks or functions. A module or component used herein may exist on a hardware component independently of other modules / components, or a module / component may include other modules / components, programs Or an element or process shared with the device. Modules or components may reside on one device, such as on a client or server, or modules / components may be distributed across multiple devices, such as on multiple clients or server machines. Any system described may be implemented on the server in whole or in part, or as part of a network service. Alternatively, systems such as those described herein may be implemented in whole or in part on a local computer or terminal. In any case, the implementation of the system provided in this application includes memory, processor, and network resources (data ports and signal lines (such as optical or electrical)) unless otherwise specified. ) May be required.

  The embodiments described herein generally require the use of a computer that includes processing and memory resources. For example, the system described herein may be implemented on a server or network service. These servers may be connected and used by a user on a network such as the Internet, or by a combination of networks such as a cellular network and the Internet. Instead, one or more embodiments described herein may be implemented locally, in whole or in part, on a computing device such as a desktop, mobile phone, personal digital assistance, or laptop computer. May be. Thus, any memory, processing, and network resources may be used to implement, use, or implement any of the embodiments described herein (including any method implementation or any system implementation). It may be used in the context.

  Further, one or more embodiments described herein may be implemented using instructions that are executable by one or more processors. These instructions may be executed on a computer readable medium. The apparatus shown in the following drawings provides examples of processing resources and computer readable media on which instructions that implement embodiments of the present invention may be stored and / or executed. Specifically, the multiple devices shown in the embodiments of the present invention include one or more processors and various forms of memory for holding data and instructions. Examples of computer readable media include a permanent memory storage device such as a hard drive on a personal computer or server. Other examples of computer storage media include CD or DVD units, portable storage units such as flash memory (such as those provided on many cell phones or personal digital assistants (PDAs)), and magnetic memory. . Computers, terminals, and network-enabled devices (eg, mobile devices such as mobile phones) are all examples of machines and devices that utilize instructions stored on a processor, memory, and computer-readable media.

1 illustrates a system for identifying user genre and fashion style preferences using visual information according to one or more embodiments. 6 illustrates a method for predicting user preferences for a particular genre according to one or more embodiments. FIG. 6 illustrates an example of a panel that can be generated to present a set of visual materials to a user to request the user to provide a response in one embodiment. FIG. 6 illustrates a panel that allows a user to select size information for various types of fashion products such as issues, tops, bottoms, and addresses. Fig. 3 shows a panel that allows a user to define or notify a user's preferences for unique patterns, colors, and shapes. Fig. 4 illustrates a method for predicting a genre or style of a product by a program according to an embodiment. FIG. 3 illustrates a method for matching a product according to one or more embodiments to customer preferences. FIG. 6 shows a results panel that informs a user's fashion genre preferences as determined by a program according to one embodiment. FIG. 6 illustrates a method for determining a descriptive classification and category of a fashion product provided by a fashion product content item in one or more embodiments. 1 illustrates a system for recommending fashion products to a user using product class / category determination and user activity information according to one embodiment.

Shopper's Genre and Style Preference Learning FIG. 1 illustrates a system for identifying user genre and fashion style preferences using visual information according to one or more embodiments. A system such as the one shown in FIG. 1 is an image of a pre-selected fashion product to determine likes, dislikes, preferences, and other user feedback to confirm a user's style or genre preferences. To people. In contrast to the described embodiments, conventional techniques for estimating a shopper's (e.g., user or customer) style or genre preference typically best represents their personal preference for style and genre. It involves steps to ask people questions about the genre / style. However, there are problems with conventional methods, one of which is that language is not always suitable for capturing fashion preferences and expressions. Furthermore, users do not always know what their preferences are.

  Accordingly, the embodiment described herein and with reference to FIG. 1 includes a system that programmatically learns a user's fashion style and genre preferences using visual materials and photographs. System 100 may be provided in a variety of computing environments including client-server architectures. For example, the system 100 is implemented on one or more servers (or other computing devices) to provide services such as those illustrated by one or more embodiments detailed herein. May be. In this environment, the system 100 may be implemented on a website such as an e-commerce site, a search engine, or a shopping portal.

  The system 100 may rely on a genre definition defined by an expert or operator. For example, fashion genres include (but are not limited to) “looks” with genres of chic, street, Boho, urban / hip hop, and conservative. For example, an expert may select clothing and clothing ensembles that represent various categories (the number of which is set by design or choice). In some cases, such representative apparel and apparel ensembles can be used in (i) determining a user's genre preference and (ii) predicting the genre of another apparel or apparel item. It forms the data, the point of comparison. In one embodiment, the system 100 shows images of clothing and clothing ensembles when worn. For example, images of people wearing different genres of clothing and accessories (including celebrity images) can be presented to the user. Responding to the individual images allows the user to define whether the presented clothing is in the user's preferred style or type. Thus, the system can learn from user selections performed on images rather than text explanations about preferences and user self-reports.

  More particularly, the system 100 includes a user interface 110, a user database 120, a genre score component 130, a genre terminator 134, a visual material component 140, and a product database 150. A user of the system 100 may correspond to a fashion product shopper or customer. In some embodiments, system 100 is implemented on an online medium. For example, system 100 is implemented as part of an e-commerce site, shopping portal, or other web-based or networked environment that gives people the opportunity to browse (and potentially purchase) fashion products. Is possible. The interface 110 may correspond to, for example, a web page or interactive functionality provided on the web page.

  A user of system 100 is associated with a profile in user database 120. For example, a user may make a transaction with a service provider that provides the system 100. Alternatively, the user may be perceived by cookie / computer information, by account / login, or in a separate online session with a provider of system 100 (eg, an e-commerce site). Regardless of the determination of the system 100 and the user's style or genre preferences, the user may interact with the interface 110 and may provide parameters 112 related to the fashion product that the user can wear. Parameters that may be defined by the user include, for example, the user's clothing size, preferred price ranges for fashion and clothing items, and preferred brand names. The user may also declare information about clothing and apparel visual characteristics that the user likes or dislikes. For example, a user may specify preferred colors for a particular type of clothing, preference information about fabrics or materials, preferred styles for shoes or apparel, jewelry types, and disgust or preference for particular types of patterns. .

  Product database 150 holds information from fashion product content items. As described in some embodiments, the product database 150 may store information about the fashion product shown in the product content item. Such information may be determined programmatically from analysis of images, text, and metadata of fashion product content items provided by retailers, manufacturers, and other suppliers of fashion products. Information determined by the program relating to the presented fashion product is associated in database 150 with corresponding product content items such as electronic catalog pages and sections.

  In order to programmatically determine a user's genre / style preference, the visual material component 140 is configured to present the derived image, ie, visual material, to the user to perform a genre / style preference determination. Has been. Visual material component 140 communicates fashion product visual material 152 to the user via interface 110. In one embodiment, the visual material 152 is worn (eg, worn by a celebrity or model, worn on a mannequin, or computer generated on a human image). Present fashion products or ensembles of fashion products. 2 and 3A show examples of how visual material 152 can be constructed for presentation to the user.

  Through the interface 110, the user can (i) like or dislike the user for a particular fashion product or ensemble, (ii) the user's preference in comparison to one of the fashion products, and / or (iii) the user for the fashion product. An input may be provided that indicates a score or feedback indicating the level of likes or dislikes. Visual material component 140 presents a set of visual materials 152 that require the user to enter a response that informs the user's visual preferences for the fashion genre indicated by the visual material. Further, as shown in one embodiment of FIG. 2 or 3A, the visual material may be presented to the user in the form of a quiz or game. In the form of a quiz or game, a panel is presented to the user that individually shows competing fashion products of different genres. For each fashion product, the user can respond to each panel by notifying his / her preference or dislike at the other end of the panel.

  The genre score component 130 records and determines a genre score from user input. The genre score component 130 may record responses from users presented in the visual material 152 to score individual fashion genre categories. Optionally, the genre score component 130 may also record and score user responses to fashion genre subcategories in combination with the visual material component 140.

  A number of techniques may be used to identify a user's fashion genre preferences. In one embodiment, a set of visual materials 152 is pre-designed to present several images of fashion products for each identified genre. A user simply responds with a preference or liking / dislike input when viewing an image of a fashion product to notify his / her liking or liking in comparison to another of one fashion genre. The genre score component 130 holds a genre score 133 indicating the user's genre preference for the genre represented by the set of visual materials 152. The genre score 133 can be recorded in the user database 120 in relation to the profile from the user.

  In addition, or alternatively, the genre terminator 134 may determine the user's one or more preferred genres / or subcategories (eg, first, second, and third genres) based at least in part on the score 133. ). In addition, the genre terminator 134 and the score 133 may use information based on a probabilistic assumption that a user who has a preference for a specific genre is likely to have a specific liking or dislike of another genre. In respect, the visual material 152 output for the user by the component 140 may be influenced. For example, a user with a business genre preference may be considered less likely to prefer street genre clothing.

  Further, in one or more embodiments, the genre descriptor 151 is tagged to the fashion product identified in the product database 150. These descriptors include genre descriptors determined by the program, which can be determined by the product genre predictor component (PGPC) 154. Specifically, the PGPC 154 analyzes the product content item to obtain information that can be used to determine one or more genres of fashion products indicated in the content item. Thus, the system 100 can be used to determine the user's genre preference and to predict the category and category of fashion product genres. The determined genre descriptor 151 from the PGPC may include a sub-genre or genre category that includes determination of the second and third genres. For example, many fashion products may share more than one genre. FIG. 4 illustrates a method for predicting one or more genres of fashion products using fashion product content items according to some embodiments. As shown, the determination may be based on a learning operation using a ground torus product set 155 provided by an operator of the system 100.

  In one embodiment, system 100 also includes a product recommendation engine 170. According to one or more embodiments, the product recommendation engine 170 can: (i) user information identifying genre / style preferences and parameters of fashion products that the user can purchase; and (ii) fashion product information. Based on that, recommend fashion products to users. User information 172 is provided from the user database 120. Specifically, user information 172 is provided by genre preference and / or genre preference information 137 output by genre score component 130. The fashion product information 174 is acquired from the product database 150. The fashion product information 174 includes genre classifications and / or subcategories associated with individual products predicted by the program. The fashion product information 174 may also include information obtained from the fashion product content item and a tag (eg, metadata) provided by the fashion product content item or the supplier of the underlying fashion product. . Based on the user information and the fashion product information, the recommendation engine 170 can recommend individual fashion products among the products identified in the product database 150, for example. The recommended product 176 may notify the user via the interface 110.

  Furthermore, in an embodiment, the system 100 can show its confidence regarding the prediction of the user's genre and style. As described below, the system 100 also allows the user to record known parameters such as the user's clothing size, price preferences, and / or his / her own likes / dislikes for a particular style, pattern, and color. Includes interface. This information is used in matching products to user preferences. The system generally includes a first or top level genre that predicts a broad range of genres or style matching, a second or second level genre that predicts multiple fine-grained genres and styles, multiple Allows multiple levels of genre prediction, such as a third or third level genre that predicts domain-specific styles.

  FIG. 2 illustrates a method for predicting user preferences for a particular genre according to one or more embodiments. More particularly, the method as described determines, for a particular user, the first, second, and third genres of that user's preferences. The method as described may be implemented using a system such as that shown in FIG. Accordingly, reference is made to the elements and reference numbers in FIG. 1 for suitable elements and components for performing the described steps or sub-steps.

  A set of images is presented to the user (210). In one embodiment, the visual material generator 140 selects individual images in the set and displays them to the user via the user interface 110. The set of images can be selected in advance so as to be composed of various genres. In one embodiment, some or all of the genres have been determined using manual definition and selection. Accordingly, manual input may be used to sort the image sets into different genres and classify each image into a specific genre. Instead, some or all of the images in the set are determined by the program to be associated with a genre. For example, programmatic methods may be used to identify similarities between items of clothing, and similarity comparisons may be used to associate clothing with a particular genre.

  The user is requested to respond by responding to the individual presentation of each image and providing input (via interface 110) that informs the user whether they like or dislike the image. The user response is recorded (220). In one implementation, this input is one of the games that the user can participate in (“hot-or-knot game”) with an input that indicates whether the user considered the individual images in the set as hot-or-not. As a part, it is required for users.

  Based on the user's input, the user's genre preference is determined (230). Referring to one embodiment of FIG. 1, a genre terminator 134 determines user preferences for a genre. In one embodiment, genre terminator 134 uses an algorithm to determine a user's genre preference (eg, first, second, and third). In one embodiment, the algorithm is used as follows.

The entire set of genres is expressed as S = {S _i , i = 1... N}, where n is the number of genres. Assume that each user has a pre-determined preferred set of genres denoted F = {F _j , j = 1... M}. Also assume that for each genre S _i a set of images denoted T _i is provided. Here, when image pairs (I ₁ , I ₂ ) are presented to the user from different image sets T _i and T _j , (i) when only one of these belongs to F (Assuming it is I ₁ without losing generality), the user selects I ₁ with a higher probability of p> 0.5, and (b) both images belong to F If none of the images belong to F, it is assumed that the user randomly selects any image with a probability of 0.5.

After each round r, the algorithm holds a vector of user probability estimates denoted Q _r = [q _r1 , q _r2 ,... Q _m ] belonging to each genre. After each style of question in a round, the algorithm will update the user's genre probability. This update can be performed as follows. That is, of the two genres presented to the user, the one selected by the user is updated using the following equation.

q _{r + 1} = q _r * p (1)
Anything not selected by the user is updated using the following formula:

q _{r + 1} = q _r * (1-p) (2)
All remaining genres are updated using the following formula:

q _{r + 1} = q _r * 0.5 (3)
After the update, normalize Q _{r + 1} so that the sum of all probabilities is equal to 1.

  Based on the current genre probability, do one of the following:

  1. If one of the genre probabilities is above a certain threshold, the algorithm ends the test and returns the best genre to the user.

  2. If none of the genre probabilities exceeds the threshold, the algorithm selects two genre images to present to the user in the next round.

  (I) a method of randomly selecting two genres and / or (ii) selecting the top two genres with the highest probabilities (this method is useful for rapidly converging to a guess with correct probability. Two images can be selected for the next round of testing using different schemes, such as minimizing the number of image pairs presented to the user.

  The algorithm can be generalized and k (k> 2) images can be presented to the user. In this case, the update formulas of (2) and (3) will be as follows.

_{qr + 1} = _qr * (1-p) / (k-1) (4)
q _{r + 1} = q _r * 1 / k (5)
It is also possible to generalize the algorithm and determine t (t> 1) genres. In this case, the criterion for stopping can be changed and the top t probabilities can be checked. The method for selecting the next set of images requires selecting images from both the top t genres and the rest of the genres.

  According to one embodiment, the user's response to likes and dislikes notifications is used to determine the user's favorite first, second, and third genres (240). The favorite first, second and third genres can be determined simultaneously. One way to do this is to predict the first, second, and third genres in order. However, an approximation algorithm can be used to minimize the number of images presented to the user (and thereby reduce the number of user responses). When all of the images used for the prediction of the first genre are tagged with both the second and third genres, the image selected by the user in the prediction of the first genre is used, and the image for the second genre is used. Multiple histograms can be constructed, one and multiple for the third genre (one per domain). The highest genre in these histograms can be used to predict the second and third genres.

In order to provide a good user experience, some embodiments provide progress feedback, notifying the amount of progress the user has made toward computer learning in the user's preferred genre. In one embodiment, a progress bar can be presented to the user to notify the progress of genre prediction. The distance max q _ri between the threshold and the current best genre probability can be used as an indicator of progress.

  FIG. 3A illustrates an example of a panel that can be generated to present visual material 152 (FIG. 1) to the user to request the user to provide a response in one embodiment. In FIG. 3A, panel 310 is presented through interface 110 (see FIG. 1). Thus, for example, the panel 310 can be formatted as a web page. Panel 310 has image pairs 312, 314, each showing a particular genre of clothing (weared by a celebrity or model). The user selects one image in comparison with another and informs his / her preference for the particular genre presented by that image (in comparison with the genre presented by the other image) Is possible. Thus, the user's selection of one image in comparison with another is an input that informs the user's preference in comparison with another in one genre. Once the user has made a selection, the visuals component 140 presents the user with another panel having different image pairs (representing different genres of clothing) to ask the user for a similar selection. . According to one embodiment, the comparison game between image pairs can be continued over several rounds, and the user's choice / dislike of various genres defined by the system 100, depending on the user's choice in each round. Information about

  FIG. 3B shows a panel 330 that allows a user to select size information for various types of fashion products, such as issues, tops, bottoms, and addresses. Using parameters such as size, it is possible to recommend fashion products, filter recommendations for users based on the availability of a given size, or the user's genre It is also possible to modify the preferences to accommodate the user's specific size or body type.

  In addition to recording user feedback on genre selection (with competing clothing images), in some embodiments, the user improves or extends genre determination with inputs that define some user preferences. It is also possible. FIG. 3C shows a panel 350 that allows the user to define or notify the user's preferences for unique patterns, colors, and shapes. Properties that allow the user to define preferences may be specific to a particular type of fashion product. For example, user-definable shape preferences can be presented as specific to shoes, or more specifically to the category of fashion products for women's shoes.

Associates and ensemble recommendations for clothing, apparel, and accessories According to an embodiment, an online commerce environment (such as that implemented by the system of FIG. 1) recommends additional clothing, apparel, or accessories. A recommended engine is implemented. Such a recommendation may be made, for example, to provide a fashion ensemble or matched clothing / apparel pair.

  To facilitate apparel / apparel or accessory recommendations, in one or more embodiments, at least some products for sale are analyzed programmatically to predict the genre and style of individual products. Yes. FIG. 4 illustrates a method for predicting a product genre or style programmatically in one embodiment.

  In product genre prediction, several different feature types such as metadata features (based on text descriptions) and visual features (based on visual vocabularies computed from thousands of images) Is synthesized.

  In extracting program features, different forms of features are available 410 for individual products in the catalog. Feature extraction includes metadata extraction (414) and visual feature extraction (418). In the extraction of metadata features, the metadata features are identified and represented as vectors, but in this case appear in one of the metadata fields (title, description, brand, price, etc.). Each word or word pair represents one dimension in the vector. Visual features can be determined using image analysis and can be represented as vectors. Here, a vector can represent one global feature computed over the entire image or a visual vocabulary computed over thousands of images. These visual features include color, shape, and / or texture. The final feature vector can be computed by combining metadata and visual vectors, for example, by directly connecting metadata features and visual features to form a single large feature vector V.

  A fashion expert manually tags a set of products with first, second, and third genre tags to form a ground torus set (420).

  A machine learning algorithm (support vector machine or boosting or Bayesian learning) is used to learn 430 the mapping of these products to different genres from the extracted feature vectors. If the feature vector V is given, it is possible to train a binary classifier for each genre and determine the probability of whether the product belongs to that genre.

  Then, genre prediction can be performed for individual products not included in the ground torus set (440). For each product, the probabilities for all genres are estimated, and the highest genre is selected as the genre prediction for that product.

Multi-level level classification can be performed to estimate all of the first (444), second (446), and third genres (448) of products, in which case the second or third genre Are conditioned based on the first genre. Train the first genre classifier as described above. Then, given a first genre g _{1 of} products, for each first genre g ₁ , train a new set of second genre g ₂ and third genre g ₃ . When the feature vector P (g ₁ g ₂ g ₃ | V) is given during the test, the joint probabilities of the first and second / third genres can be calculated by the following equations.

P (g ₁ g ₂ g ₃ | V) = P (g ₁ | V) * P (g ₂ | g ₁ V) * P (g ₃ | g ₁ V) (6)
Product Matching According to the embodiment, (i) identify the predicted product genre of the product (as shown in FIG. 4) and (ii) (as shown in one embodiment of FIG. 2) Identify product recommendations by identifying the user's genre or style preference for a given apparel and apparel and matching the product to the user using (iii) (i) and (ii) Run.

  In some embodiments, products are recommended for recommendation by raising products that match user preferences to higher ranks and lowering products that do not match user preferences to lower ranks. It can be advertised intensively.

  As described herein, reducing the rank of a product that does not match user preferences may either (i) either filter out the unmatched product from the presentation to the user and / or (ii) match This can be done by lowering the unfinished product towards the end of the result. Users are matched products by regular automatic email (for example, sending email every day, twice a week, once a week, or once a month) or by logging on to a website (Or recommended products) can be viewed. The system also continues to learn the user's overall genre preferences and domain-specific genre preferences depending on how often the product is presented to the user and how often the user clicks on the product.

  More particularly, FIG. 5 illustrates a method for matching a product according to one or more embodiments to customer preferences. Refer to the components of FIG. 1 for suitable components for performing the described steps or sub-steps.

  The first and second / third genre combinations with the highest joint probabilities can be selected as product genres.

  For a given user, the user's first, second, and third genres are identified (510). For example, the results of the process as shown in FIG. 2 may be analyzed and searched to determine the user's favorite genre. Visual material component 140 may request the user to present and respond to visual material 152. A series of requests may be presented to the user to allow the user to define relative preferences for various different genres. The resulting score (determined from the user response) is used to determine the user's fashion genre preference.

  Once the first, second, and third genres of user preferences are identified, a group of products matching the user preferences is identified from the product database 150 (520).

  In one embodiment, a selection, filtering and weighting process is applied 530 to the matching products to identify a subset of fashion products recommended to the user. For example, selection and filtering may be performed to eliminate fashion products that are not available and not the user's size or that have a color, pattern, or shape that the user defines as disliked. . As another example, items with colors, brands, or keywords that are not the user's preference may be removed by filtering matching products. Also, weighting matching products may favor / avoid fashion products that meet the user's defined preferences for color, pattern, shape, or brand, for example. As a result, matching products can be presented to the user, for example, as a search or browse list (540). In one embodiment, the remaining products are then sorted by matching score to determine the order in which they are presented to the user.

  According to one embodiment, the matching score can be computed as a linear combination of different individual matching scores.

s = w ₁ * a ₁ + w ₂ * a ₂ +. . . Here, (w _i > 0).

  Individual matching scores include probabilities for the first, second, and third genres of products, age matching scores, price preferences, and other color, style, or pattern preferences.

Presenting results Various process, algorithm, and system output results can be provided to the user in various forms, but some embodiments can be used by the user to determine the user's fashion genre preferences Interactive tools. FIG. 6 shows a result panel that informs the user of the fashion genre preferences as determined by the program according to one embodiment. Results panel 610 can be output in response to individual participation in quizzes or questions generated through visual component 140, for example. Through a process as shown in the various embodiments, the results panel 610 can be used to display one of the user's first genre (sporty) and an additional second genre (conservative) or third genre (modern, boho). May be identified. The results panel 610 may also display fashion products that match the user's genre / style preference. An image of a fashion product may be preselected based on an image that is considered to represent a particular genre or combination of genres. Alternatively, some or all of the fashion products shown may be selected for the user. For example, parameters such as color preferences defined by the user may be used to present several items of clothing or apparel. Similarly, if the user prefers a particular style of shoe (eg, a boot defined by the user via an interface such as that shown in FIG. 3C), the boot generates a footwear result panel 610. Also good.

Improved Feature Representation The embodiments described herein may include an improved feature representation of a fashion product descriptive classification. Specifically, the description classification can be defined so that an operator (for example, an expert) includes a plurality of categories (or sub-classifications). According to embodiments, fashion product content items (eg, clothing catalogs or web images) are analyzed and features are extracted from images, text, and metadata. The extracted features are then analyzed to associate the fashion product with one or more descriptive classifications (of the fashion product) and associate each of the one or more categories with the descriptive classification.

  FIG. 7 illustrates a method for determining a description classification and category of a fashion product provided by a fashion product content item in one or more embodiments.

  The operator defines the description classification and category (or sub-classification) of the fashion product (710). In one embodiment, the descriptive classification includes (but is not limited to) genre, shape or silhouette, pattern, and color. For example, the following classification may be used.

Genre tag (GT): Silhouette tag such as classic, trendy, edgy, boho, etc. (ST): Pattern tag (PT) such as high heel, open toe, sleeveless, turtleneck, etc. Color family tag (CFT) such as floral print, stripe, zebra print, etc. : For each fashion product content item, such as bright red, dark pink, etc., extract a set of basic visual and text features from the content item (720). These features include, for example, color histograms, shape descriptors, texture features, and text description features. Image recognition and text analysis (including text metadata analysis) can be performed on individual content items to determine such characteristics.

An analysis is performed on the primitive features to determine the classification and categorization (or sub-classification) of the product indicated in the content item (730). This analysis can be quantitative. More particularly, in one embodiment, the analysis can be statistical. Furthermore, multiple methods can be implemented to associate fashion products with classifications. For color classification, generate a set of cluster centers based on manually ground labeled ground truth. Each product (or its image) is assigned to the nearest cluster based on its distance in the histogram space.

f is a primitive feature vector containing visual and text information,
c _i (where i = 1,... N ^CFT ) is the cluster centroid of the color family / classification,
X _i ^CFT is a component of the color family ^{hyperdimension} X ^CFT ,
g is a mapping from distance to likelihood.

A support vector machine classifier (SVM) may be used to associate or assign products to the classification. For each category, an operator (eg, a fashion expert) has a set of positive examples with characteristics corresponding to the tag and a set of negative examples that do not have the characteristics corresponding to the tag. And select. With the appearance of a new (unknown) item, a trained SVM is used to generate a decision value from the item's visual and textual features. This decision value represents the distance of the item to the separation hyperplane. Only the value on the positive side of the hyperplane is retained.

Here, α _i ^T , c _i ^T , and b _i ^T are learned SVM parameters corresponding to the respective tags i of the respective superdimensions Tε {GT, ST, PT}. As before, f is the basic feature vector of the item and g _j is all other items in the training set. K is a kernel function and T (x) = xH (x), where H (x) is a snake side function.

  FIG. 8 illustrates a system for recommending fashion products to users using product class / category determination and user activity information according to one embodiment. A system as shown in the embodiment of FIG. 8 may represent a modification or variation to the embodiment shown in FIG. 1 and elsewhere in the application. Accordingly, the functions and components of FIG. 8 may optionally be viewed as a complement or extension of the system such as that shown in FIG.

  The system 800 may include a user database 120 and a product database 150. As described in other embodiments, the user database 120 associates specific information with individual users, such as the user's fashion genre preferences (which may be determined by the program) and parameters defined by the user. In addition, the user database 120 may be coupled to a monitor component 810 that monitors or detects any user activity and related activities related to fashion product content items. May be. The monitoring component 810 can: (i) user interaction with a search result including a user's selection of a particular item in the search result or otherwise notice of interest in it; and (ii) an online browsing or shopping environment; Activity such as one or more of user interactions between and the like may be detected. Information 812 identifying the item of interest (eg, product) can be stored in the database 120. In one embodiment, this information 812 includes items that are displayed to the user and clicked by the user, and items that are displayed to the user but not clicked.

  User monitor 810 may detect session-specific activity or activity history 814 in the user's past sessions. The history of activity can extend to search conditions that the user has entered, for example, at a search engine or e-commerce site. User interactions may be detected through interface 810 or through browser or browser data (eg, browser history and cookie information). In some embodiments, the activity history 814 includes queries entered by the user, impressions (ie, items retrieved by the search engine and presented to the user), and purchase clicks (ie, clicked by the user). Item). Project the query set onto the aforementioned fashion-oriented feature space, thereby obtaining some positive training samples.

  In one embodiment, product database 150 is coupled to product category / class terminator 820. The category / class terminator 820 may analyze the fashion product content item to determine one or more classifications / categories 822 for each product. In one embodiment, the category / class terminator 820 performs a process such as that shown in FIG. In one embodiment, the resulting descriptive classification / categorization is stored in the product database 150.

  According to one or more embodiments, user preference profiler 830 generates user profile 832 based on activity information 812 and / or history information 814. The profiler 830 updates the user profile 832 for each individual user. In generating and updating the user profile 832, the profiler 830 (i) the fashion product from the user's activity information 812 (for example, the product the user selected to browse during browsing or searching, the user did not select the view) Products), (ii) determine the classification and categorization of those products using the product database 150 (as shown in FIG. 7), and (iii) are identified from the activity information 812 The user's profile 832 is generated using the product description classification and categorization. User profile 832 may identify, augment, or otherwise identify a user's fashion genre preferences. Therefore, the user profile 832 may be combined with the determination of the fashion genre by the program shown in the other embodiments, or may be used as an alternative to this. For example, user profile 832 is session-specific to determine that the user is looking for event-specific clothing (eg, evening gown) that would otherwise not be in the user's preferred genre. And may be stable. The profiler 830 may generate the profile 832 using the history information 814.

  In one embodiment, the recommendation engine 170 selects for the user based at least in part on the genre preference identified by the user profile 832 and / or the genre preference identified via the visual / score component. Configured to recommend data for the product 176 that has been selected. The recommendation engine 170 may include history data 814 as a component for determining the recommended product 176. The recommendation engine 170 may also be used to recommend and / or search and / or re-rank products in response to a user query / search or request for a particular type of fashion product.

Short-term usage Embodiments recognize that in online scenarios, a user's short-term preferences can be important. Furthermore, embodiments also recognize the need for an online algorithm that learns quickly from user activity and that immediately improves the user's shopping and search experience. For example, when a user is shopping for a formal holiday party or resort vacation, the user's long-term preferences, such as color, pattern, and brand, are not much to improve the overall shopping experience. It will not be useful. Therefore, a system that learns about the user in real time when the user is interacting with the site can achieve a more appropriate result.

In one embodiment, on the online system, when a user performs a query and performs a click, these contents are stored in a daily user profile. A preference summary is generated by kernel density estimation and retained for use in ranking. As the user enters a query and clicks on item i, a feature vector indicating the characteristics of item i is fetched (from a precomputed table) and the kernel density estimator is updated online. Thus, the data is efficiently aggregated in the daily user profile generation model.

Where n is the number of clicks on the user's session and h is the kernel bandwidth. Using the function p, it is possible to score the relevance to the current session of the feature vector x ^T of the material. A secondary kernel may be used. After the user enters a query, all items related to the query (relevance based on visual and text) are fetched from the item database along with the corresponding absolute ranking. Then, the score of each retrieved item is calculated according to the above-described offline and online models.

CONCLUSION While a number of embodiments have been described in terms of fashion products, alternative embodiments may extend to various types of products. Specifically, embodiments may extend to other products that are generally classified by personal taste and appearance, such as furniture, carpets (and curtains), and design exteriors.

CONCLUSION While exemplary embodiments have been described in detail herein with reference to the accompanying drawings, it is to be understood that the described embodiments are not limited to the specific examples described above. Thus, many modifications and variations are possible, including matching a feature described in one embodiment to another embodiment that does not refer to that feature. In addition, specific features described individually or as part of one embodiment may be described separately, even if other features and embodiments do not refer to that particular feature. It can also be combined with some of the features or other embodiments.

Claims (18)

A computer-implemented method for determining user preferences for fashion products,
Using one or more processors,
Programmatically determining a user's fashion preferences based on user interactions with a plurality of fashion product content items individually indicating corresponding fashion products;
Making recommendations for fashion products to the user based at least in part on the user's fashion preferences;
A computer-implemented method comprising the steps of: Individually displaying the plurality of fashion product content items to the user;
Requesting the user for responses indicating likes and dislikes of the plurality of fashion product content items;
The computer-implemented method of claim 1, further comprising: The step of determining the fashion preference by a program includes:
Identifying a set of images that individually show one or more fashion items;
Displaying a sequence including a plurality of panels to the user, each panel including at least two images from the set;
For each panel, recording a response from the user indicating the user's most preferred or disliked of the at least two images in the panel;
The computer-implemented method of claim 1, comprising:   The step of displaying the sequence displays a corresponding fashion product in which each fashion product content item of the individual panel has a corresponding genre that is different from the fashion product of the other fashion product content item of the panel. The computer-implemented method of claim 3, further comprising the step of generating each panel.   The computer-implemented method of claim 4, wherein the fashion product content item of each panel is determined to belong to the corresponding genre by manual input.   The computer-implemented method of claim 1, further comprising requesting the user to provide an input that defines one or more known parameters relating to the user's fashion preferences.   The computer-implemented method of claim 1, wherein the one or more known parameters include a size or price preference of the user.   The step of recommending the fashion product to a user includes recommending one or more fashion products based on the determined fashion preference and known parameters of the user. The computer-implemented method according to 1. A computer-implemented method using a fashion product program descriptor,
Using one or more processors,
Analyzing the fashion product content item and determining a set of features of the fashion product indicated in the fashion product content item;
Programmatically associating the fashion product with a predefined descriptive category for each of a plurality of descriptive categories based on the quantitative analysis of the determined set of features;
Determining or predicting user preferences using the product content item and a predefined description category for each of the plurality of description categories;
A method comprising the steps of:   The method of claim 9, wherein the plurality of description classifications includes one or more of a genre class, a pattern class, a shape class, or a color family class.   The method of claim 9, wherein analyzing a fashion product content item comprises performing image analysis on an image portion of the fashion product content item.   The step of programmatically associating the fashion product with the predefined descriptive category determines a probability that the fashion product has a visual characteristic of each of the predefined categories of the descriptive classification. The method of claim 9 including the step of:   Using the fashion product content item and a predefined description category for each of the plurality of description categories detects a user selection or interaction with the fashion product content item; and The method of claim 9, comprising using the predefined description category of each of the description classifications to determine the user preference.   The step of detecting a user's selection or interaction with the fashion product content item comprises determining the user's profile based on the predefined description category of the individual description category of each product viewed by the user. The method of claim 13, comprising monitoring a fashion product content item that the user has selected to view for determination.   Using the fashion product content item and a predefined descriptive category for each of the plurality of description categories identifies a user's fashion genre or style preference, and the fashion product content item The method of claim 9, comprising recommending or not recommending the fashion product based on the pre-defined description category associated with the user's fashion genre or style preference. Recording historical information related to the user ’s online activity on fashion products;
Determining a preference of the user's genre for a fashion product based at least in part on the history information;
10. The method of claim 9, further comprising: A computer-implemented method for determining user preferences for fashion products,
Using one or more processors,
Analyzing individual fashion product content items representing a catalog of fashion products and determining, for each fashion product content item, a set of features of the fashion product indicated in the fashion product content item;
Based on a quantitative analysis of the determined set of features, each fashion product represented by one of the fashion product content items is assigned a predefined description category for each of a plurality of description categories. Step by step programmatically,
Detecting one or more fashion product content items that may be of interest to the user;
Using the predefined descriptive categories for each of the plurality of descriptive categories of the one or more fashion product content items considered interesting to the user to determine the user's fashion preferences;
A method comprising the steps of:   The method of claim 7, wherein determining the user preference includes using historical information including search conditions previously used by the user.

Images