A product recommendation method, device and computer-readable storage medium

Abstract

The present invention discloses a product recommendation method, device and computer-readable storage medium, the method comprising: using the product rating to obtain the recommendation prediction score of the product to be recommended, using user data to obtain the user's purchase probability of the product category; combining the recommendation prediction score of the product to be recommended and the purchase probability of the product category to obtain the recommended product. In this way, the present invention can accurately recommend products that meet the actual needs of the user.

Description

Technical Field

[0001] The present invention relates to the field of data mining technology, and in particular to a commodity recommendation method, device and computer-readable storage medium. Background Art

[0002] At present, with the development of e-commerce, more and more consumers choose online shopping platforms to shop. In order to increase the purchase rate of consumers, various e-commerce platforms usually display recommended products that consumers are interested in on some pages. Therefore, the more the product recommendation method is close to the interests of consumers, the more effective it is in promoting shopping.

[0003] Currently, collaborative filtering based on products is widely used in recommendation systems. However, this traditional model relies on users' ratings of purchased products when constructing the user-product rating matrix. In most cases, the number of products purchased by users is limited, resulting in a sparse user-product rating matrix, which may lead to biased recommendation results. In addition, recommendations based on item similarity only consider user preferences, but do not actually measure user purchasing needs, and the evaluation factors for generating recommendations to users are relatively limited. Summary of the invention

[0004] The main technical problem solved by the present invention is to provide a commodity recommendation method, which can accurately recommend commodities that meet the actual needs of users.

[0005] In order to solve the above technical problems, a technical solution adopted by the present invention is: to provide a product recommendation method, which includes: using the product rating to obtain the recommendation prediction score of the product to be recommended, and using user data to obtain the user's purchase probability of the product category; combining the recommendation prediction score of the product to be recommended and the purchase probability of the product category to obtain the recommended product.

[0006] Among them, using the ratings of the products to obtain the recommendation prediction scores of the products to be recommended includes: obtaining the ratings of the products to obtain a rating matrix; using the ratings of the products to calculate the associations between the products; and using the rating matrix and the associations between the products to obtain the recommendation prediction scores of the products to be recommended.

[0007] Among them, the user data includes user behavior data, and obtaining the ratings of the products to obtain the rating matrix includes: obtaining the ratings of the evaluated products; combining the ratings of the evaluated products and the user behavior data to predict the ratings of the unevaluated products; and obtaining the rating matrix by combining the ratings of the evaluated products and the ratings of the unevaluated products.

[0008] Among them, combining the scores of evaluated products and the user behavior data, predicting the scores of unevaluated products includes: obtaining feature data of the unevaluated products; inputting the feature data of the unevaluated products into a product score prediction model to obtain the scores of the unevaluated products, and the product score prediction model is trained using the feature data of evaluated products and the scores of the evaluated products; wherein the features of the products include derived features, and the derived features are derived using the basic features of the products and the user behavior data.

[0009] Among them, the derived features include the user's preference for the product, and the presentation form of the user's preference for the product includes one or more of the similarity between the product and the product purchased by the user, the matching degree between the product and the user's characteristics, and the user's attention to the product.

[0010] Among them, obtaining the user's preference for the product includes: taking a weighted average of at least two presentation forms to obtain the user's preference score for the product; calculating the proportion of each product's preference score exceeding similar products to obtain the user's preference for the product.

[0011] Among them, taking weighted average of at least two presentation forms includes: the weight of each presentation form is the product of the inverse of the evaluation period and the influence value, the influence value is the influence of the presentation form on the product score, the influence of the presentation form on the product score includes the proportion of positive reviews in the product score under the presentation form, and the evaluation period is the data collection period.

[0012] Among them, the similarity between the product and the product purchased by the user includes one or more of the similarity ratio between the product characteristics and the common characteristics of the product purchased by the user, and the number of overlaps between the product label and the label of the product purchased by the user; the matching degree between the product and the user characteristics includes one or more of the number of overlaps between the product label and the user label, and the degree of conformity between the product price range and the user's purchasing power; the user's attention to the product includes one or more of the number of searches, browses, purchases, favorites, and shares for the product by the user.

[0013] The derived features include the user's sensitivity to product characteristics, and the presentation form of the user's sensitivity to product characteristics includes one or more of the user's sensitivity to spatial characteristics of the product and the user's sensitivity to temporal characteristics of the product.

[0014] Calculating the user's sensitivity to the spatial / temporal characteristics of a product includes: obtaining the user's negative review rate for the spatial / temporal characteristics of the product in different spaces / times; calculating the proportion of the negative review rate of each space / time exceeding that of other spaces / times, and obtaining the user's sensitivity to the spatial / temporal characteristics of the product.

[0015] Among them, the derived characteristics include the matching degree between product characteristics and user demand characteristics, and the presentation form of the matching degree between product characteristics and user demand characteristics includes one or more of the probability that the product characteristics and user demand characteristics are antonymous, and the probability that the product characteristics and user demand characteristics are similar.

[0016] Among them, calculating the probability that the product characteristics are antonymous to the user demand characteristics includes: obtaining the reverse negative review rate of the product, the reverse negative review rate is the negative review rate of users for the antonymous features of the product, and the antonymous features are the features antonymous to the user demand characteristics; calculating the proportion of the reverse negative review rate of each product exceeding that of similar products, and obtaining the probability that the product characteristics are antonymous to the user demand characteristics; calculating the probability that the product characteristics are similar to the user demand characteristics includes: obtaining the positive positive review rate of the product, the positive positive review rate is the positive review rate of users for the positive features of the product, and the positive features are the features similar to the user demand characteristics; calculating the proportion of the positive positive review rate of each product exceeding that of similar products, and obtaining the probability that the product characteristics are similar to the user demand characteristics.

[0017] Among them, the product recommendation method also includes: replacing the missing features with the average feature value of the features corresponding to the missing features of similar users; and / or normalizing the weights of the features of products with abnormal prices / purchase quantities / purchase frequencies; and / or encoding discrete feature data; and / or performing nonlinear expansion such as polynomial expansion on some feature data.

[0018] The product rating prediction model includes any one of a regression model, a Lasso model, an elastic network model, a stepwise regression model, a robust regression model, an Xgboost model, an Adaboost model, and a random forest model.

[0019] Among them, using user data to obtain the user's purchase probability of a product category includes: obtaining the user's feature data; inputting the user's feature data into a category purchase probability prediction model to obtain the purchase probability of the product category, and the product category purchase probability prediction model is obtained by supervised training of the user's feature data.

[0020] Among them, the user's characteristic data includes the user's demand for product categories. The presentation forms of the user's demand for product categories include implicit demand and explicit demand. The user's demand for product categories is the weighted average of the implicit demand and the explicit demand. The implicit demand includes one or more of the correlation between the product category and the product purchased by the user and the correlation between the product category and the user's family structure. The explicit demand includes the user's attention to the product category.

[0021] Among them, the user's characteristic data includes: the user's family structure data; and / or the degree of match between the price range of the commodity category and the user's purchasing power; and / or the user's attention to the commodity category; and / or the user's purchase cycle for the commodity category, and whether the current time is within the purchase cycle; and / or the degree of correlation between the commodity category and the commodity purchased by the user.

[0022] The supervised model training of the user's feature data includes: marking the user's feature data, and obtaining the purchase probability of the commodity category corresponding to the user's feature data.

[0023] Among them, the product recommendation method also includes: normalizing the weights of the features of products with abnormal prices / purchase quantities / purchase frequencies; and / or encoding discrete feature data; and / or performing nonlinear expansion such as polynomial expansion on some feature data.

[0024] Among them, the category purchase probability prediction model includes any one of a regression model, a Lasso model, an elastic network model, a stepwise regression model, a robust regression model, an Xgboost model, an Adaboost model, and a random forest model.

[0025] Among them, user data includes user behavior data and user attribute information; user behavior data includes data generated by users' operational activities based on communication network conditions, and user behavior data includes on-site data and off-site data. On-site data includes users' activity behavior data on shopping websites, and off-site data includes usage behavior data of objects / Internet devices used by users.

[0026] In order to solve the above technical problem, another technical solution adopted by the present invention is: to provide a product recommendation device, the product recommendation device includes a processor, and the processor is used to execute instructions to implement the above product recommendation method.

[0027] In order to solve the above technical problem, another technical solution adopted by the present invention is: providing a computer-readable storage medium, which is used to store instructions / program data, and the instructions / program data can be executed to implement the above-mentioned product recommendation method.

[0028] The beneficial effect of the present invention is that, different from the prior art, the present invention can accurately recommend to users products that meet their purchase needs and suit their tastes by respectively obtaining the recommendation prediction scores of the products to be recommended and the user's purchase probability of the product categories, and comprehensively considering the user's product preferences and purchase needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flowchart of the first product recommendation method in the implementation mode of the present application;

[0030] Figure 2 is a flow chart of the second product recommendation method in the implementation mode of the present application;

[0031] Figure 3 is a schematic diagram of a scoring matrix in an embodiment of the present application;

[0032] Figure 4 is a schematic diagram of a commodity similarity matrix in an implementation manner of the present application;

[0033] Figure 5 It is a flowchart of a product rating prediction model training method in an embodiment of the present application;

[0034] Figure 6 is a flowchart of the third product recommendation method in the implementation mode of the present application;

[0035] Figure 7 is a flowchart of the fourth product recommendation method in the implementation mode of the present application;

[0036] Figure 8 It is a flowchart of a method for training a category purchase probability prediction model in an embodiment of the present application;

[0037] Fig. 9 is a flowchart of the fifth product recommendation method in the implementation mode of the present application;

[0038] Fig.10 is a flowchart of the sixth product recommendation method in the application implementation mode;

[0039] Fig.11 It is a structural schematic diagram of a commodity recommendation device in an embodiment of the present application;

[0040] Fig.12 It is a structural schematic diagram of a commodity recommendation device in an embodiment of the present application;

[0041] Fig.13 It is a schematic diagram of the structure of a computer-readable storage medium in an embodiment of the present application. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail below with reference to the accompanying drawings and examples.

[0043] The present application provides a product recommendation method, which can accurately recommend products that meet the actual needs of users by respectively obtaining the recommendation prediction score of the recommended product and the user's purchase probability of the product category, and comprehensively considering the user's product preferences and purchase needs.

[0044] See also Figure 1 , Figure 1is a flowchart of the first product recommendation method in the embodiment of the present application. It should be noted that if there are substantially the same results, this embodiment does not Figure 1 The process sequence shown is limited. Figure 1 As shown, this embodiment includes:

[0045] S110: Using the ratings of the products, obtain the recommendation prediction scores of the products to be recommended, and using the user data, obtain the purchase probability of the user for the product category.

[0046] S130: Combining the recommendation prediction score of the product to be recommended and the purchase probability of the product category to obtain a recommended product.

[0047] Online shopping platforms offer a wide variety of products. For ease of management, products can be classified into a variety of categories, such as clothing, home appliances, and daily necessities. Furthermore, product categories can be divided into more detailed categories, such as clothing includes tops, pants, shoes, hats, etc., home appliances include TVs, refrigerators, washing machines, air conditioners, etc., and daily necessities include tables, chairs, toothpaste, toothbrushes, shampoos, etc. Users can choose to buy products in different categories according to their needs. After purchasing a product, users can rate the product according to their satisfaction. Different people may have different satisfaction levels with the same product, so the same product can have multiple ratings.

[0048] Since the rating of a product can reflect the user's satisfaction with the product, the product that the user may be interested in can be inferred based on the user's rating of the product, so as to recommend it to the user. For example, based on the user's liking for product S, similar / related products to S can be recommended to the user. There can be multiple products to be recommended, and the products to be recommended can be scored and given a recommendation prediction score to distinguish the degree of recommendation of the products and select the most recommended products. For example, the higher the recommendation prediction score, the higher the corresponding recommendation degree, that is, the higher the recommendation prediction score, the more likely it is to be recommended to the user. Of course, the scoring rules for the recommendation prediction score are not limited here and can be set according to needs.

[0049] The products recommended based on product ratings are, to a certain extent, what the user is interested in, but they are not necessarily what the user wants to buy. In other words, the recommended products may not meet the user's purchasing needs and cannot better improve the purchase rate.

[0050] In this implementation, based on the product rating, the user's purchase needs are also considered, and the two are combined to recommend the products that the user wants to buy. Among them, due to the variety of products, the amount of calculation to calculate the purchase probability of each product separately will be very large. In this application, the purchase probability of the product category is calculated, which can reflect the user's purchase probability of the products in a certain category to a certain extent, and more products under this category can be recommended, such as preferentially recommending products that the user is interested in under this category. That is, the recommendation prediction score of the recommended product and the purchase probability of the product category are combined, and the product with a high recommendation prediction score under the product category with a high purchase probability is selected as the recommended product, and the recommended product is recommended to the user.

[0051] In this implementation, by respectively obtaining the recommendation prediction score of the recommended product and the user's purchase probability of the product category, and integrating the user's product preferences and purchase needs, it is possible to accurately recommend products that meet the user's purchase needs and suit their tastes.

[0052] See also Figure 2 , Figure 2 is a flow chart of the second product recommendation method in the embodiment of the present application. It should be noted that if there are substantially the same results, this embodiment does not Figure 2 The process sequence shown is limited. In this implementation, based on the collaborative filtering principle of commodities, commodity recommendations can be made according to commodity ratings. Figure 2 As shown, this embodiment includes:

[0053] S210: Obtain ratings of products and obtain a rating matrix.

[0054] See also Figure 3 , Figure 3 Schematic diagram of a rating matrix in the implementation of the present application. By combining the ratings of multiple products, a user-product rating matrix A can be obtained. The matrix includes multiple ratings of multiple products and the user corresponding to each rating. If a user does not take any action on a product, that is, the user does not search, browse, add to cart, collect, share, purchase but not rate the product, etc., then the corresponding element value in the matrix is ​​set to 0. For example, user a scores product 101 as 5 and does not take any action on product 104.

[0055] S230: Calculate the correlation between products using the ratings of the products.

[0056] See also Figure 4 , Figure 4It is a schematic diagram of a commodity similarity matrix in the implementation mode of the present application. The similarity between each commodity is calculated using the user-commodity rating matrix A to obtain the commodity similarity matrix B. Among them, the similarity between commodities is a quantitative representation of the degree of association between commodities, and cannot be simply understood as the similarity of the surface features / functions of commodities. That is, this method may calculate that the washing machine is associated with the refrigerator. For example, the similarity between commodity 103 and commodity 104 is 3.

[0057] S250: Obtain the recommendation prediction score of the product to be recommended by using the rating matrix and the correlation between the products.

[0058] Multiply the user-item rating matrix A by the item similarity matrix B to obtain the user-item recommendation prediction score matrix C. The elements in the matrix represent the recommendation prediction scores of the items to be recommended. For different users, the same item can have multiple recommendation prediction scores, and items can be recommended to users based on the recommendation prediction scores.

[0059] In this implementation, based on the collaborative filtering principle of products, product recommendations are made according to product ratings. This can effectively utilize similar information of other users based on the preferences in the user ratings, discover the user's potential but undiscovered interests and preferences, and recommend products that meet the user's preferences to the user.

[0060] However, when shopping, users may not rate all purchased items, and the number of items purchased by users is limited, which can easily cause the user-item rating matrix to be sparse, resulting in the possibility of biased recommendation effects.

[0061] In this application, the ratings of unevaluated products can be predicted, so that unevaluated products also have ratings. Combined with the ratings of evaluated products, the amount of data in the original rating matrix of collaborative filtering can be increased, which is more conducive to calculating product similarity, increasing the number of products to be recommended and the accuracy of the predicted scores of recommended products. Unrated products include products that users have searched, browsed, added to cart, collected, shared, purchased but not rated, and other activities.

[0062] In one embodiment, the ratings of reviewed products and user behavior data may be combined to predict the ratings of unevaluated products.

[0063] User data includes user behavior data and user attribute information; user behavior data includes data generated by users' operations based on communication network conditions. The so-called communication network conditions refer to the behavior data that users perform through the network, such as browsing websites, using apps, using Bluetooth to transmit data, etc., which can be electronically recorded and stored.

[0064] User behavior data includes on-site data and off-site data. On-site data includes user activity behavior data on shopping websites, and off-site data includes user usage behavior data of physical / Internet devices.

[0065] The data on the site includes the user's activity data and related product information in the mall. Obtain the user's ID in the mall and the corresponding tag under the ID, including tags of the characteristics of the user's purchased products and tags of the user's purchasing power; obtain the user's attribute characteristics, including the user's age, gender, geographic location, etc.; obtain the user's search, browsing, adding to cart, collection, sharing, purchase, rating and other action footprints on the product.

[0066] Off-site data includes data on products used by users at home and actual usage data of the products. Obtain the ID of the products connected to the Internet in the user's home, information about the product ID, and tags under the product ID. The information about the product ID includes the weight, volume, and energy efficiency rating of the product. Tags under the product ID include the price of the product, the category to which the product belongs, the main and secondary tags of the product, and the spatial attributes of the product. Obtain the time and functions of the products connected to the Internet in the user's home.

[0067] Off-site data also includes users’ reading data on web pages other than shopping websites. Reading data includes reading content that users browse or collect in online content recommendations that include categories that are not bound to the Internet, and product purchase tips that users browse or collect in content recommendations.

[0068] In order to predict the rating of unevaluated products, the present application provides a product rating prediction model, which is obtained by supervised training using the features and ratings of evaluated products. The features of the evaluated products can be used as input, and the ratings of the evaluated products can be used as output to train the model and obtain the product rating prediction model.

[0069] See also Figure 5 , Figure 5 is a flowchart of the product rating prediction model training method in the embodiment of the present application. It should be noted that if there are substantially the same results, this embodiment does not Figure 5 The process sequence shown is limited. Figure 5 As shown, this embodiment includes:

[0070] S510: Acquire feature data of the evaluated product.

[0071] The characteristics of a product include the basic characteristics of the product, such as the volume, weight parameters, performance indicators and other product characteristics that can be directly obtained. Specifically, it may include the product ID, product ID information, labels under the product ID, etc. The product ID information is the physical properties of the product, including the weight, volume, energy efficiency level, etc. of the product. The labels under the product ID include the price of the product, the category to which the product belongs, the main and secondary labels of the product, the spatial attributes of the product, etc. For example, a product is a dishwasher, the price of the product is 4,000-6,000, the category of the product is a dishwasher, the spatial attribute of the product is kitchen, and the electrical attribute of the product is hands-free.

[0072] The characteristics of a product also include its derived characteristics, which are derived from the basic characteristics of the product and user behavior data, such as the user's preference for the product, the user's satisfaction with the product's performance characteristics, the probability that the product matches the user's needs, and other product characteristics that need to be obtained indirectly. In other words, characteristics with user factors can be derived. There can be many types of derived characteristics depending on the needs.

[0073] In one embodiment, the derived features include the user's preference for the product, especially the user's preference for a certain product among similar products. There may be multiple presentation forms to reflect the user's preference for a certain product. Such as the similarity between the product and the product purchased by the user, the matching degree between the product and the user's characteristics, the user's attention to the product, etc. The similarity between the product and the product purchased by the user may be the similarity ratio between the product characteristics and the common characteristics of the product purchased by the user, the number of overlaps between the product label and the label of the product purchased by the user, etc. The matching degree between the product and the user's characteristics may be the number of overlaps between the product label and the user label, the degree of conformity between the product price range and the user's purchasing power, etc. The user's attention to the product may be the number of searches, browsing, purchases, collections, and sharing of the product by the user. Certain weights may be assigned to these presentation forms, and one or more of these presentation forms may be weighted averaged to calculate the user's preference for the product. For example, the following calculation formula may be used to calculate the user's preference score for the product:

[0074] Preference score = (w 1 *Number of times the product label overlaps with the user's purchased product label + w 2 *Number of times the product label and user label overlap + w 3 * Number of searches + w 4 *Views+w 5 *Number of additional purchases + w 6 *Collection times+w 7 *The price range of the product) / (w 1 +w 2 +w 3 +w 4 +w5 +w 6 +w 7 ).

[0075] Among them, data within a certain period of time can be collected to calculate the preference score. Each evaluation can be called an evaluation cycle. The weight (W) of each presentation form is the product of the inverse of the evaluation cycle and the influence value. The influence value is the degree of influence of the presentation form on the product score. The degree of influence of the presentation form on the product score includes the proportion of good reviews in the product score under the presentation form. If it has been analyzed that the proportion of good reviews in the user's collection is k, then k is used as the influence value of the collection of this presentation form. Calculate the proportion of the preference score of the product that exceeds similar products, and use this proportion as the user's preference for the product. For example, there are 20 models of wall-mounted air conditioners. The user preference score of a certain product exceeds 89% of the products, then 89% is used as the preference degree of the product.

[0076] In one embodiment, the derived features include the user's sensitivity to the product characteristics, and the user's sensitivity to the product characteristics is presented in the form of the user's sensitivity to the spatial characteristics of the product, the user's sensitivity to the temporal characteristics of the product, etc. Because users have higher requirements for some characteristics of the product in a specific space or a specific time, such as users have higher requirements for product noise control in the bedroom at 10 pm than in the living room at 7 pm, that is, they are more sensitive to noise and are more likely to give bad reviews. Therefore, the bad review rate of products related to user sensitive factors belonging to a specific space / time is obtained, and the bad review rate is used as the user's sensitivity value to the space / time characteristics. Specifically, the bad review rate of users to the space / time characteristics of the same product in different spaces / times is obtained; the proportion of the bad review rate of each space / time exceeding other spaces / times is calculated to obtain the user's sensitivity to the product characteristics.

[0077] In one embodiment, the derived features include the degree of matching between the product characteristics and the user demand characteristics, and the presentation of the degree of matching between the product characteristics and the user demand characteristics includes the probability that the product characteristics are antonymous to the user demand characteristics, the probability that the product characteristics are similar to the user demand characteristics, etc. When purchasing a product, the user will consider the degree of conformity between the product label and his own needs. Obtain the proportion of negative reviews of the features antonymous to the user needs in the product reviews, calculate the average of the proportion of negative reviews of all antonymous features, and obtain the probability that the product is antonymous to the user needs; obtain the proportion of positive reviews of the features similar to the user needs in the product reviews, calculate the average of the proportion of positive reviews of all similar features, and obtain the probability that the product is similar to the user needs. Among them, calculating the probability that the product characteristics are antonymous to the user demand characteristics includes: obtaining the reverse negative review rate of the product, the reverse negative review rate is the negative review rate of the user to the antonymous features of the product, and the antonymous features are the features opposite to the user demand characteristics; calculate the proportion of the reverse negative review rate of each product that exceeds similar products, and obtain the probability that the product characteristics are antonymous to the user demand characteristics. Calculating the probability that product characteristics are similar to user demand characteristics includes: obtaining the positive praise rate of the product, the positive praise rate is the rate of users' praise for the positive features of the product, and the positive features are features similar to user demand characteristics; calculating the proportion of each product's positive praise rate that exceeds similar products, and obtaining the probability that product characteristics are similar to user demand characteristics.

[0078] In one embodiment, the derived features include the probability that the product matches the user's needs. The user's family environment and demographic structure will affect the user's choice of products. The degree of matching between the physical attributes of the product's features and the user's needs is obtained to obtain the probability that the product matches the user's needs. For example, when the family space is large, when choosing the same model of air conditioner, the user will prefer a larger air conditioner.

[0079] In one embodiment, the derived features include the average score of the product features and the space features to which they belong. The features of the product are divided into main features and secondary features, and the space features to which they belong are divided into main features and secondary features, and the average scores of the product and the space to which they belong under multiple main features and multiple secondary features are obtained respectively. When purchasing a product, users will not only consider the demand characteristics of the product itself, but also the demand characteristics of the environment in which the product is used. For example, a product is an air conditioner, and the main features of this model of air conditioner are rapid cooling and heating, and the secondary features are energy saving and power saving. The average scores of this model of air conditioner under rapid cooling and heating and energy saving and power saving are obtained respectively; the space to which the air conditioner belongs is a bedroom, and the main features of the bedroom are silent, and the secondary features are windless, and the average scores of this model of air conditioner under silent and windless are obtained respectively.

[0080] In one embodiment, the derived features include the probability that the product features match the product features of the category in the reading content browsed or collected by the user.

[0081] S530: Process the feature data of the evaluated products to obtain sample feature values.

[0082] The sample feature value is the feature value of the sample used for training. The basic features and derived features of the above-mentioned products include continuous feature data and discrete feature data; some derived features can only be obtained by user operations, and not all products of all users can obtain all the above derived features. Therefore, it is necessary to further process the basic features and derived features of the products to make better use of these feature data.

[0083] When some features of some users are missing, the average feature value of similar users can be used instead. Obtain data such as the user's own label attributes and the label attributes of the purchased goods, use cluster analysis to calculate the distance value between users, select the m users closest to the target user as similar users, and use the average feature value of the m similar users under this feature to replace the feature value of the target user under this feature. When the target user is a new user and lacks behavioral data, and similar users cannot be calculated, use the average feature value of the target user's region under this feature to replace the feature value of the target user under this feature.

[0084] When some feature data are abnormal, the data weights are normalized. For example, during promotional holidays, there is a lot of user behavior data on products. At this time, the price influence plays a dominant role, but the actual product rating will not be higher because of the low price. At this time, the weights of the product data and user behavior data of the promotional holiday need to be normalized. In other words, the same user buying the same product in large quantities multiple times can be called abnormal data.

[0085] When some features are discrete data, the data needs to be quantized. In this embodiment, the discrete feature data is encoded. When some features are associated with each other, the interaction between the features is considered, and some feature data is nonlinearly expanded such as polynomial.

[0086] The above-mentioned commodities have many basic and derived features, and not all of them can be well utilized, so it is necessary to process and select the feature data. For example, the high-order features use the PCA dimensionality reduction method to extract the main feature components in the features; or the Lasso model is used to perform L1 regularization on the feature data to compress the data structure; or the recursive elimination method, integrated model, etc. are used to select multiple feature data with better model combination effects to obtain sample feature values. In this implementation, the method of processing and selecting the feature data is not limited.

[0087] S550: Taking the feature value of the evaluated product as input and the rating of the evaluated product as output, the model is trained to obtain a product rating prediction model.

[0088] In this implementation, multiple models such as generalized regression models and integrated models can be used for comprehensive evaluation. Commonly used generalized regression models include regression models, Lasso models, elastic network models, stepwise regression models, robust regression models, etc. Commonly used integrated models include Xgboost models, Adaboost models, random forest models, etc. The feature values ​​and scores of the evaluated products are input into the model, and the model is trained in a supervised manner, wherein a model with low prediction accuracy and prediction deviation is selected as the product rating prediction model.

[0089] In this implementation, by analyzing the features of the evaluated products and combining them with user behavior data to extract features that have a high degree of influence on user ratings, the product rating prediction model is trained. This can better utilize existing user data, use the user's real rating data to train the model, and increase the accuracy of the model's prediction results.

[0090] See also Figure 6 , Figure 6 is a flow chart of the third product recommendation method in the embodiment of the present application. It should be noted that if there are substantially the same results, this embodiment does not Figure 6 The process sequence shown is limited. Figure 6 As shown, this embodiment includes:

[0091] S610: Predicting the rating of the unevaluated product using the product rating prediction model.

[0092] Obtain the basic features of the unevaluated product, and use the basic features of the unevaluated product and user behavior data to obtain the derived features of the unevaluated product. Process and select the basic features and derived features of the unevaluated product to obtain the feature values ​​for prediction of the unevaluated product. Please refer to the above description for the method of obtaining the basic features and derived features of the unevaluated product, which will not be repeated here. Input the predicted feature values ​​into the trained product rating prediction model, extract the feature vector of the unevaluated product, and predict the user's rating of the unevaluated product.

[0093] S630: Combining the ratings of the evaluated products with the ratings of the unevaluated products to obtain a rating matrix.

[0094] The rating matrix of the evaluated products of multiple users is obtained, and the ratings of the unevaluated products are added to the rating matrix of the evaluated products to obtain the user-product rating matrix A.

[0095] S650: Calculate the correlation between products using the product ratings.

[0096] The user-item rating matrix A is used to calculate the similarity between each item and obtain the item similarity matrix B.

[0097] S670: Obtain the recommendation prediction score of the product to be recommended by using the rating matrix and the association relationship between the products.

[0098] Multiply the user-item rating matrix A by the item similarity matrix B to obtain the user-item recommendation prediction score matrix C. The elements in the matrix represent the user's recommendation prediction score for the recommended item. Items can be recommended to users based on the prediction score.

[0099] In this implementation, a product rating prediction model is used to predict the ratings of unevaluated products, so that the ratings of unevaluated products are more consistent with the actual ratings of users. At the same time, the data volume of the original rating matrix of collaborative filtering is increased, the possibility of default value assignment is improved, the matrix sparsity problem is solved, and it is more conducive to calculating product similarity. The probability of the unreasonable preferences of a few people affecting the model is reduced, and the number of products to be recommended and the accuracy of the recommended prediction scores of the products to be recommended are increased.

[0100] See also Figure 7 , Figure 7 is a flowchart of the fourth product recommendation method in the embodiment of the present application. It should be noted that if there are substantially the same results, this embodiment does not Figure 7 The process sequence shown is limited. Figure 7 As shown, this embodiment includes:

[0101] S710: Obtain user feature data.

[0102] User feature data includes user data and user-derived feature data. User-derived feature data is derived from user data, such as user's family structure, user's demand for category goods, user's purchase cycle for category goods, and other user feature data that needs to be obtained indirectly.

[0103] In one embodiment, the user-derived feature data includes the user's family structure. The user's family structure specifically includes the number of people in the user's family, the family population structure, the home space, etc. The number of people in the user's family can be judged by the volume, size and other physical attributes of the home appliances purchased by the user, the functional attributes or the functional use of the networked goods, such as using the water level used for cooking to judge the number of people eating. The population structure of the elderly and children in the family can be judged by the category attributes of the goods bound to the network and the functional attributes used. For example, if the user purchases a children's air conditioner and uses functions such as children's anti-kicking blankets, it is judged that there are children in the user's family. If the user creates and uses the elderly-related functional modes in the APP applicable to the network, it is judged that there are elderly people in the user's family. The size of the furniture space can also be comprehensively evaluated through home appliances such as the living room, kitchen, and bedroom, such as using the size of the air conditioner in the living room / bedroom to predict the area of ​​the living room / bedroom, using the size of the living room TV to predict the area of ​​the living room, and using the volume of the refrigerator to judge the kitchen area. It can also be actively provided by the user when registering.

[0104] In one embodiment, the user-derived characteristic data includes the user's demand for category goods. The user's demand for category goods includes implicit demand and explicit demand, and the demand for the category is obtained by combining implicit demand and explicit demand. Implicit demand includes demand inferred from family space, associated equipment, family structure, etc., such as judging the categories needed by the family based on the family space, and finding the categories and quantities of home appliances that the user already has in the home appliance linkage, and comprehensively judging that the user has implicit demand for the category of goods that need to be continuously supplemented; such as judging that the user has implicit demand for the supplementary product category of the product based on the networked product, such as the networked device is an air conditioner, then the user is judged to have an implicit demand for the air conditioner companion; if the user browses or collects the purchase or use of the non-networked product category in the network content recommendation, then the user is judged to have implicit demand for the product category. Explicit demand includes demand obtained according to the user's attention level, obtaining the user's search, browsing, adding to cart, collection, sharing, purchase and other action footprints for the category goods, and performing weighted average on each action to calculate the user's attention level to the category goods. The weight corresponding to the action is determined according to the time interval from the time the action occurs to the present, and the weight is the inverse of the exponential function of the time interval. The closer the action occurs to the present, the greater the corresponding weight, and the greater the weight becomes, that is, the closer the action occurs to the present, the greater the user's attention to the product in this category. The demand degree of this category is calculated by weighting the implicit demand and explicit demand. The demand degree calculation formula of this category is as follows:

[0105] Demand level = (w 1 *Implicit requirement + w 2 Explicit demand) / (w 1 +w 2 );

[0106] Among them, the frequency of user purchases in different branch categories is calculated as the weight of implicit demand and explicit demand.

[0107] In one embodiment, the user-derived characteristic data includes the user's purchase cycle for category goods. The purchase cycle of the category goods is to determine whether the current time is within the purchase cycle range of the category. For general goods, after a user purchases once, he will purchase again after a period of consumption. According to the user's previous purchase data rules, it is determined whether the current time is within the user's purchase cycle.

[0108] In one embodiment, the user-derived feature data includes the probability of the user repurchasing the category goods. For example, home appliances are not only related to rigid needs, but also to decoration needs. The study of user purchasing behavior found that if it is found that within a certain period of time, the user has the possibility of purchasing a set of products, and some users have the possibility of purchasing the whole house needs, then the frequency of repurchasing the whole house home improvement category should also be increased. The probability of the user repurchasing the category goods is calculated by the number of users with this feature / the number of features with this behavior.

[0109] In one embodiment, the user-derived feature data includes the price range of the category of the goods purchased by the user. The demand frequency of goods of the same category at different prices is different. The price range of the category of the goods purchased by the user is obtained by comprehensively analyzing the actual price of the goods in the category and the price estimated by different market groups in the planning.

[0110] In one embodiment, the user-derived feature data includes the user's behavior data on the purchased category goods. The number of times and the time interval between the user's search, browsing, adding to cart, collection, sharing, purchase and other actions on the purchased category goods within a period of time from now are calculated.

[0111] The above user data and user-derived feature data are further processed to make better use of these feature data. Abnormal data are weighted normalized; discrete feature data are encoded; some feature data are nonlinearly expanded such as polynomials; the above user data and user-derived feature data are processed using one or more models such as PCA dimensionality reduction method, Lasso model, recursive elimination method, and integrated model to obtain model input feature values. The specific data processing method is the same as the data processing method for the basic features and derived features of the product mentioned above, and will not be repeated here.

[0112] S730: Input the user's feature data into a product category purchase probability prediction model to obtain the purchase probability of the product category.

[0113] In one embodiment, the category purchase probability prediction model is trained using previous user features.

[0114] See also Figure 8 , Figure 8 is a flow chart of the method for training a category purchase probability prediction model in the embodiment of the present application. It should be noted that if there are substantially the same results, this embodiment is not based on Figure 8 The process sequence shown is limited. Figure 8 As shown, this embodiment includes:

[0115] S810: Acquire previous user characteristic data.

[0116] The previous user feature data is the user feature data during model training, and the user feature data includes user data and user-derived feature data.

[0117] S830: Process the previous user feature data to obtain the user input feature value.

[0118] The user data and user derived feature data are further processed to make better use of these feature data. The specific data processing method is the same as the data processing method of the basic features and derived features of the above-mentioned goods, and will not be repeated here.

[0119] S850: Train a category purchase probability prediction model using user input feature values ​​of previous user features.

[0120] In this implementation, multiple models such as generalized regression models and integrated models can be used for comprehensive evaluation. Commonly used generalized regression models include regression models, Lasso models, elastic network models, stepwise regression models, robust regression models, etc. Commonly used integrated models include Xgboost models, Adaboost models, random forest models, etc. The user input feature values ​​of previous user features and manually set probability labels are input into the model, and the model is supervised and trained, wherein a model with low prediction accuracy and prediction deviation is selected as the category purchase probability prediction model.

[0121] In this implementation, the user's demand for goods in different categories is taken into consideration, and the user's previous data is used to train the purchase probability prediction model to increase the accuracy of the model's prediction. The category purchase probability prediction model is used to predict the user's current purchase probability for goods in each category, thereby overcoming the deficiency of the traditional model that it can only recommend preferred goods.

[0122] See also Fig. 9 , Fig. 9 is a flowchart of the fifth product recommendation method in the embodiment of the present application. It should be noted that if there are substantially the same results, this embodiment does not Fig. 9 The process sequence shown is limited. Fig. 9 As shown, this embodiment includes:

[0123] S910: Obtain user behavior data and ratings of reviewed products.

[0124] S930: Combine the ratings of the reviewed products and the user behavior data to predict the ratings of the unevaluated products.

[0125] S950: Obtain a rating matrix by combining the ratings of the rated products and the ratings of the unevaluated products.

[0126] S970: Calculate the correlation between products using the product ratings.

[0127] S990: Obtain the recommendation prediction score of the product to be recommended by using the rating matrix and the correlation between the products.

[0128] In this implementation, the ratings of reviewed products and user behavior data are combined to predict the ratings of unevaluated products, so that the ratings of unevaluated products are more consistent with the actual ratings of users. At the same time, the amount of data in the original rating matrix of collaborative filtering is increased, the possibility of assigning default values ​​is improved, the matrix sparsity problem is solved, and it is more conducive to calculating the similarity of products. The probability of the unreasonable preferences of a few people affecting the model is reduced, and the number of products to be recommended and the accuracy of the recommended prediction scores of products to be recommended are increased.

[0129] See also Fig.10 , Fig.10 is a flowchart of the sixth product recommendation method in the embodiment of the present application. It should be noted that if there are substantially the same results, this embodiment does not Fig.10 The process sequence shown is limited. Fig.10 As shown, this embodiment includes:

[0130] S1010: Obtain user behavior data and ratings of reviewed products.

[0131] S1020: Combine the ratings of the reviewed products and the user behavior data to predict the ratings of the unevaluated products.

[0132] S1030: Obtain a rating matrix by combining the ratings of the rated products and the ratings of the unevaluated products.

[0133] S1040: Calculate the correlation between products using the product ratings.

[0134] S1050: Obtain the recommendation prediction score of the product to be recommended by using the rating matrix and the correlation between the products.

[0135] S1060: Utilize user data to obtain the user's purchase probability for a product category.

[0136] S1070: Combining the recommendation prediction score of the product to be recommended and the purchase probability of the product category to obtain a recommended product.

[0137] Based on the recommendation prediction score of the product to be recommended and the purchase probability of the product category to which it belongs, the product category with a high purchase probability and a high recommendation prediction score is recommended to the user.

[0138] In one embodiment, the recommendation prediction scores of the recommended products and the purchase probabilities of the product categories are arranged in descending order, and the product categories with purchase probabilities greater than a first threshold are selected as recommended product categories, and the recommended products with recommendation prediction scores greater than a second threshold and whose product categories belong to the recommended product categories are selected as recommended products.

[0139] In one embodiment, the recommendation prediction scores of the recommended products and the purchase probabilities of the product categories are arranged in descending order, and the product categories with purchase probabilities greater than a first threshold are selected as the recommended product categories, and the top N recommended products that belong to the recommended product categories are selected as the recommended products.

[0140] In one embodiment, the purchase probabilities of commodity categories are arranged in descending order, commodity categories with a purchase probability greater than a first threshold are selected as recommended commodity categories, commodities belonging to the recommended commodity categories are selected from the commodities to be recommended, and the commodity categories are arranged in descending order according to the recommendation prediction scores, and in each category, commodities to be recommended with a recommendation prediction score greater than a second threshold are selected as recommended commodities or the top N commodities to be recommended are selected as recommended commodities.

[0141] In this implementation, by using a rating prediction model to rate products that users have not rated, the possibility of assigning default values ​​is increased, the matrix sparsity problem is solved, the probability of the unreasonable preferences of a few people affecting the model is reduced, and the number of products to be recommended and the accuracy of the recommendation prediction scores of the products to be recommended are increased; at the same time, a category purchase probability prediction model is used to solve the deficiency of the traditional model that only preferred products can be recommended; combined with the user's product preferences and purchasing needs, it can accurately recommend to the user products that have purchase needs and meet their tastes.

[0142] See also Fig.11 , Fig.11 1 is a schematic diagram of the structure of a commodity recommendation device in an embodiment of the present application. In this embodiment, the commodity recommendation device includes an acquisition module 111 and a recommendation module 112.

[0143] The acquisition module 111 is used to obtain multiple filter parameters of multiple filters in the neural network to be compressed; the analysis module 112 is used to obtain the recommendation prediction score of the recommended product using the product rating, and obtain the user's purchase probability of the product category using the user data; the recommendation module 112 is used to combine the recommendation prediction score of the recommended product and the purchase probability of the product category to obtain the recommended product. The product recommendation device is used to obtain the recommendation prediction score of the recommended product and the user's purchase probability of the product category, respectively, and comprehensively consider the user's product preferences and purchase needs, so as to accurately recommend products that meet the user's actual needs.

[0144] See also Fig.12 , Fig.12 1 is a schematic diagram of the structure of a commodity recommendation device in an embodiment of the present application. In this embodiment, the commodity recommendation device 121 includes a processor 122 .

[0145] The processor 122 may also be referred to as a CPU (Central Processing Unit). The processor 122 may be an integrated circuit chip having signal processing capabilities. The processor 122 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. A general-purpose processor may be a microprocessor or the processor 122 may also be any conventional processor, etc.

[0146] The product recommendation device 121 may further include a memory (not shown in the figure) for storing instructions and data required for the processor 122 to run.

[0147] The processor 122 is used to execute instructions to implement the method provided by any embodiment of the product recommendation method of the present application and any non-conflicting combination.

[0148] See also Fig.13 , Fig.13 It is a schematic diagram of the structure of the computer-readable storage medium in the implementation mode of the present application. The computer-readable storage medium 11 of the embodiment of the present application stores instructions / program data 12, and when the instructions / program data 12 are executed, the method provided by any embodiment of the product recommendation method of the present application and any non-conflicting combination is implemented. Among them, the instructions / program data 12 can form a program file and be stored in the above-mentioned storage medium 11 in the form of a software product, so that a computer device (which can be a personal computer, server, or network device, etc.) or a processor (processor) executes all or part of the steps of each implementation method of the present application. The aforementioned storage medium 11 includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0149] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0150] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0151] The above description is only an implementation mode of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A product recommendation method, characterized in that: The method comprises: The recommendation prediction score of the product to be recommended is obtained by using the product rating, and the purchase probability of the user for the product category is obtained by using the user data, wherein the product rating includes the rating of the evaluated product and the rating of the unevaluated product, and the rating of the unevaluated product is obtained by combining the rating of the evaluated product and the user behavior data prediction; Combining the recommendation prediction score of the to-be-recommended product and the purchase probability of the product category to obtain a recommended product; The step of combining the rating of the evaluated product and the user behavior data to predict the rating of the unevaluated product includes: Acquiring characteristic data of the unevaluated product; Inputting the feature data of the unevaluated product into a product rating prediction model to obtain a rating of the unevaluated product, wherein the product rating prediction model is trained using the feature data of the evaluated product and the rating of the evaluated product; Among them, the characteristics of the product include derived characteristics, which are derived from the basic characteristics of the product and the user behavior data. The derived characteristics include the user's sensitivity to the product characteristics. The presentation form of the user's sensitivity to the product characteristics includes one or more of the user's sensitivity to the spatial characteristics of the product and the user's sensitivity to the temporal characteristics of the product.

2. The commodity recommendation method according to claim 1, characterized in that: The step of obtaining the recommendation prediction score of the to-be-recommended product by using the rating of the product includes: Obtain the ratings of the products and obtain a rating matrix; Calculating the association between products using the ratings of the products; The recommendation prediction score of the to-be-recommended product is obtained by utilizing the correlation relationship between the rating matrix and the products.

3. The commodity recommendation method according to claim 2, characterized in that: The user data includes user behavior data, and the step of obtaining the rating of the product to obtain the rating matrix includes: Get the ratings of reviewed products; Combining the ratings of the reviewed products and the user behavior data, predicting the ratings of unevaluated products; The rating matrix is ​​obtained by combining the ratings of the rated products and the ratings of the unevaluated products.

4. The commodity recommendation method according to claim 1, characterized in that: The derived features also include the user's preference for the product, and the presentation form of the user's preference for the product includes one or more of the similarity between the product and the product purchased by the user, the matching degree between the product and the user's features, and the user's attention to the product.

5. The commodity recommendation method according to claim 4, characterized in that: Obtaining the user's preference for a product includes: Take a weighted average of at least two presentation forms to obtain the user's preference score for the product; The proportion of the preference score of each product exceeding similar products is calculated to obtain the user's preference degree for the product.

6. The commodity recommendation method according to claim 5, characterized in that: The weighted average of at least two presentation forms comprises: The weight of each presentation form is the product of the inverse of the evaluation cycle and the influence value. The influence value is the influence of the presentation form on the product score. The influence of the presentation form on the product score includes the proportion of positive reviews in the product score under the presentation form. The evaluation cycle is the data collection cycle.

7. The commodity recommendation method according to any one of claims 4 to 6, characterized in that: The similarity between the product and the product purchased by the user includes one or more of the similarity ratio between the product features and the common features of the product purchased by the user, and the number of overlaps between the product label and the label of the product purchased by the user; The matching degree between the product and the user characteristics includes one or more of the number of overlaps between the product label and the user label, and the matching degree between the product price range and the user's purchasing power; The user's attention level to the product includes one or more of the number of searches, views, purchases, favorites, and shares of the product by the user.

8. The commodity recommendation method according to claim 1, characterized in that: Calculating the user's sensitivity to the spatial / temporal characteristics of the product includes: Obtain the negative review rate of users for spatial / temporal features of products in different spaces / times; The ratio of the negative review rate of each space / time exceeding that of other spaces / times is calculated to obtain the user's sensitivity to the space / time characteristics of the product.

9. The commodity recommendation method according to claim 1, characterized in that: The derived features also include a matching degree between product characteristics and user demand characteristics, and the presentation form of the matching degree between product characteristics and user demand characteristics includes one or more of a probability that the product characteristics and user demand characteristics are antonymous, and a probability that the product characteristics and user demand characteristics are similar.

10. The commodity recommendation method according to claim 9, characterized in that: Calculating the probability that the product characteristics are the opposite of the user demand characteristics includes: Obtain the reverse negative review rate of the product, where the reverse negative review rate is the negative review rate of users on the antonymous features of the product, where the antonymous features are features that are the opposite of the user's demand features; calculate the proportion of the reverse negative review rate of each product that exceeds similar products, and obtain the probability that the product feature is the opposite of the user's demand feature; Calculating the probability that the product characteristics are similar to the user demand characteristics includes: The positive praise rate of the product is obtained, where the positive praise rate is the praise rate of users for the positive features of the product, and the positive features are features similar to the user's demand features; the proportion of the positive praise rate of each product exceeding that of similar products is calculated, and the probability that the product characteristics are similar to the user's demand features is obtained.

11. The commodity recommendation method according to claim 1, characterized in that: The method further comprises: replacing the missing feature with the average feature value of the feature corresponding to the missing feature of similar users; and / or Normalize the weights of features for items with unusual prices / volumes / frequency; and / or Encoding discrete feature data; and / or Perform nonlinear expansion such as polynomial on some feature data.

12. The commodity recommendation method according to claim 1, characterized in that: The product rating prediction model includes any one of a regression model, a Lasso model, an elastic network model, a stepwise regression model, a robust regression model, an Xgboost model, an Adaboost model, and a random forest model.

13. The commodity recommendation method according to claim 1, characterized in that: The method of using user data to obtain the user's purchase probability of a product category includes: Obtain user feature data; The user's feature data is input into a category purchase probability prediction model to obtain the purchase probability of the commodity category, wherein the commodity category purchase probability prediction model is obtained by performing supervised training on the user's feature data.

14. The commodity recommendation method according to claim 13, characterized in that: The user's characteristic data includes the user's demand for a product category. The presentation forms of the user's demand for a product category include implicit demand and explicit demand. The user's demand for a product category is a weighted average of the implicit demand and the explicit demand. The implicit demand includes one or more of the correlation between the product category and the product purchased by the user and the correlation between the product category and the user's family structure. The explicit demand includes the user's attention to the product category.

15. The commodity recommendation method according to claim 13, characterized in that: The user's characteristic data includes: User's family structure data; and / or The degree to which the price range of the product category matches the user's purchasing power; and / or The user's interest in the product category; and / or The user's purchase cycle for the product category, and whether the current time is within the purchase cycle range; and / or The degree of correlation between the product category and the products purchased by the user.

16. The commodity recommendation method according to claim 13, characterized in that: The supervised model training of the user's feature data includes: The characteristic data of the user is marked to obtain the purchase probability of the commodity category corresponding to the characteristic data of the user.

17. The commodity recommendation method according to claim 13, characterized in that: The method further comprises: Normalize the weights of features for items with unusual prices / volumes / frequency; and / or Encoding discrete feature data; and / or Perform nonlinear expansion such as polynomial on some feature data.

18. The commodity recommendation method according to claim 13, characterized in that: The category purchase probability prediction model includes any one of a regression model, a Lasso model, an elastic network model, a stepwise regression model, a robust regression model, an Xgboost model, an Adaboost model, and a random forest model.

19. The commodity recommendation method according to claim 1, characterized in that: The user data includes user behavior data and user attribute information; the user behavior data includes data generated by the user's operating activities based on communication network conditions, and the user behavior data includes on-site data and off-site data. The on-site data includes the user's activity behavior data on the shopping website, and the off-site data includes the user's usage behavior data of objects / Internet devices.

20. A commodity recommendation device, characterized in that: The commodity recommendation device includes a processor, and the processor is used to execute instructions to implement the commodity recommendation method according to any one of claims 1-19.

21. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store instructions / program data, and the instructions / program data can be executed to implement the commodity recommendation method according to any one of claims 1-19.

Images