A content recommendation method, system and storage medium

Abstract

The present invention discloses a content recommendation method, system and storage medium, the method comprising the following steps: configuring a recall algorithm and a sorting algorithm, establishing an intention recognition model, selecting content that a user is interested in from content to be recommended by using the recall algorithm and the sorting algorithm; obtaining user characteristics and content characteristics, the user characteristics comprising product preferences and content preferences; the recall algorithm generates a list to be recommended according to the product preferences, the content preferences and the content characteristics; selecting samples, and training the intention recognition model according to the samples; the successfully trained intention recognition model derives the probability of user intention according to the user characteristics; the sorting algorithm generates a recommendation list according to the probability and the list to be recommended; through the above-mentioned manner, the present invention solves the problem that the functions of product recommendation and content recommendation cannot be integrated to allow users to obtain both the content and products they want.

Description

Technical Field

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

[0002] With the rapid development of Internet technology, how to mine and analyze massive amounts of network information has become a hot and difficult issue; recommendation systems can provide users with accurate and fast recommended products (such as goods, content, services, etc.), and have become a common interest and research hotspot in the industry and academia in the field of artificial intelligence in recent years; however, the practice of recommendation systems in the industry mainly focuses on product recommendations or content recommendations, such as Taobao's product recommendation system and Toutiao's content recommendation system; the defect of existing technologies is that they cannot integrate the functions of product recommendation and content recommendation, so that users can get both the content and products they want. Summary of the invention

[0003] The main technical problem solved by the present invention is that the functions of product recommendation and content recommendation cannot be integrated to allow users to obtain both the content and products they want.

[0004] In order to solve the above technical problems, a technical solution adopted by the present invention is to provide a content recommendation method, comprising the following steps:

[0005] Configure the recall algorithm and sorting algorithm to establish the intent recognition model;

[0006] Acquiring user characteristics and content characteristics, wherein the user characteristics include product preferences and content preferences;

[0007] The recall algorithm generates a list to be recommended according to the product preference, the content preference and the content feature;

[0008] Selecting samples, and training the intention recognition model according to the samples; the successfully trained intention recognition model derives the probability of the user's intention according to the user characteristics;

[0009] The sorting algorithm generates a recommendation list according to the probability and the list to be recommended.

[0010] As an improved solution, the recall algorithm generates a to-be-recommended list according to the product preference, the content preference and the content feature, further comprising:

[0011] The recall algorithm includes a real-time recall algorithm and a non-real-time recall algorithm;

[0012] The real-time recall algorithm calculates a first similarity between the content feature and the product preference and a second similarity between the content feature and the content preference through a collaborative filtering algorithm, and selects content to be recommended according to the first similarity and the second similarity to generate a first list to be recommended;

[0013] The non-real-time recall algorithm generates a second to-be-recommended list according to the product preference, the content preference and the content feature;

[0014] The to-be-recommended list is generated according to the first to-be-recommended list and the second to-be-recommended list.

[0015] As an improved solution, the real-time recall algorithm calculates the first similarity between the content feature and the product preference and the second similarity between the content feature and the content preference through a collaborative filtering algorithm, and selects the content to be recommended according to the first similarity and the second similarity to generate a first list to be recommended, further comprising:

[0016] The content features include product category and content category;

[0017] Calculating the first similarity between the commodity category and the commodity preference, and selecting content to be recommended according to the first similarity to generate a third list to be recommended;

[0018] calculating the second similarity between the content category and the content preference, and selecting content to be recommended according to the second similarity to generate a fourth list to be recommended;

[0019] The first list to be recommended is generated according to the third list to be recommended and the fourth list to be recommended.

[0020] As an improved solution, the non-real-time recall algorithm generates a second to-be-recommended list according to the product preference, the content preference and the content feature, further comprising:

[0021] The non-real-time recall algorithm includes a product recall algorithm and a content recall algorithm;

[0022] The content features include content quality and keywords of the content;

[0023] The product recall algorithm calculates a first recall score according to the product preference and the content quality weight; and selects content to be recommended according to the first recall score to generate a fifth list to be recommended;

[0024] The content recall algorithm generates a sixth to-be-recommended list according to the content preference and the content feature;

[0025] The second list to be recommended is generated according to the fifth list to be recommended and the sixth list to be recommended.

[0026] As an improved solution, the content recall algorithm generates a sixth to-be-recommended list according to the content preference and the content feature, further comprising:

[0027] The content preference includes content category preference and content keyword preference;

[0028] Calculating a third recall score according to the content category preference and the content quality weight, and selecting to-be-recommended content according to the third recall score to generate a seventh to-be-recommended list;

[0029] Setting a word vector of a content keyword, calculating a semantic similarity between the content keyword preference and the word vector, and selecting content to be recommended according to the semantic similarity to generate an eighth list to be recommended;

[0030] The sixth list to be recommended is generated according to the seventh list to be recommended and the eighth list to be recommended.

[0031] As an improved solution, the sample selection step further includes:

[0032] The user characteristics include activity level and target behavior;

[0033] Setting an activity level threshold; selecting a target user, wherein the activity level of the target user exceeds the activity level threshold;

[0034] The samples include positive samples and negative samples. The target user who has the target behavior is taken as a positive sample, and the target user who does not have the target behavior is taken as a negative sample.

[0035] As an improved solution, the step of training the intention recognition model according to the sample further includes:

[0036] Set accuracy threshold;

[0037] The user features of the positive sample and the negative sample are passed through the intent recognition model to generate accuracy;

[0038] If the accuracy exceeds the accuracy threshold, the intention recognition model training is successful;

[0039] If the accuracy is lower than the accuracy threshold, the intent recognition model is modified and the next training is started.

[0040] The present invention also provides a content recommendation system, comprising: a configuration unit, an acquisition unit and an execution generation unit;

[0041] The configuration unit is used to configure the recall algorithm and the sorting algorithm; and to establish an intent recognition model;

[0042] The acquisition unit is used to acquire user characteristics and content characteristics;

[0043] The execution generation unit is used to execute the recall algorithm according to the user characteristics and the content characteristics to generate a list to be recommended; to run the intention recognition model according to the user characteristics to generate the probability of user intention; and to generate a recommendation list according to the probability and the list to be recommended.

[0044] As an improved solution, the configuration unit is used to select a target user according to the user characteristics, and train the intent recognition model according to the target user.

[0045] The present invention also provides a computer storage medium for storing computer software instructions used for the above content recommendation method, which includes a program designed for executing the above content recommendation method.

[0046] The beneficial effects of the present invention are:

[0047] 1. The content recommendation method described in the present invention generates diverse preferences of users by fusing the content preference and the product preference through the recall algorithm; and trains the target users through the intention recognition model to generate the probabilities of different user intentions, so as to facilitate content recommendation based on user intentions.

[0048] 2. The content recommendation system of the present invention generates the recommendation list through the configuration unit and the execution unit according to the user characteristics and the content characteristics, so that the content recommendation is more accurate.

[0049] 3. The computer storage medium described in the present invention solves the problem of not being able to integrate the functions of product recommendation and content recommendation by executing built-in program instructions that implement the above-mentioned content recommendation method, allowing users to obtain both the content and products they want. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art are briefly introduced below; in all drawings, similar elements or parts are generally identified by similar figure marks; in the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0051] Figure 1 is a diagram of the content recommendation method architecture described in Example 1 of the present invention;

[0052] Figure 2 is a schematic diagram of the content recommendation method described in Example 1 of the present invention;

[0053] Figure 3It is a schematic diagram of the content recommendation system described in Example 2 of the present invention.

[0054] The markings of the components in the accompanying drawings are as follows:

[0055] 1- configuration unit, 2- acquisition unit, 3- execution generation unit, 100- content recommendation system. DETAILED DESCRIPTION

[0056] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0057] In the description of the present invention, it should be noted that, for example, LightGBM (Light Gradient Boosting Machine) is a framework for implementing the GBDT algorithm open sourced by Microsoft, which supports efficient parallel training; GBDT (Gradient Boosting Decision Tree) is a long-standing model in machine learning. Its main idea is to use weak classifiers (decision trees) for iterative training to generate an optimal model. This model has the advantages of good training effect and low overfitting.

[0058] In the description of the present invention, it should be noted that the terms "first", "second", "third", "fourth", "fifth", "sixth", "seventh" and "eighth" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance; in addition, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device includes a series of elements; not only those elements, but also other elements not explicitly listed, or also include elements inherent to such process, method, commodity or device. In the absence of further restrictions, the elements defined by the sentence "comprise a..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0059] In the description of the present invention, it should be noted that the reference to “embodiment” in this article means that the specific features, structures or characteristics described in conjunction with the embodiment may be included in at least one embodiment of the present invention; the appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments; those skilled in the art explicitly and implicitly understand that the embodiments described in this article may be combined with other embodiments.

[0060] Example 1

[0061] This embodiment 1 provides a content recommendation method. Figure 1 and Figure 2 , the method comprises the following steps:

[0062] In step S100, a recall algorithm and a sorting algorithm are configured to establish an intent recognition model.

[0063] In this embodiment, the recall algorithm may include multiple algorithms to achieve multi-channel recall. Multi-channel recall can not only improve the diversity of recalled content, but also enable multi-channel data to be calculated in parallel, thereby reducing system response time. The sorting algorithm is essentially a prediction model and a click-through rate estimation model. This model can predict the user's click probability for each content, that is, the user's preference for a certain content. The intent recognition model can be regarded as a classification problem, and different query intent categories are defined based on the characteristics of vertical fields. For the intent input by the user, the probability of each intent is calculated based on the statistical classification model, and the query intent is finally given.

[0064] In step S200, user characteristics and content characteristics are obtained, wherein the user characteristics include product preferences and content preferences; and the recall algorithm generates a list to be recommended according to the product preferences, the content preferences and the content characteristics.

[0065] In this embodiment, the user characteristics include: user attributes, such as gender, age, occupation, region, etc.; user behavior, such as browsing time, clicks, searches, posts, favorites, likes, etc. regardless of the window, as well as statistical indicators such as the time since user registration and activity; product preferences, such as brand, car series, price preferences, etc.; content preferences, such as content category preferences, keyword preferences, etc.; the content characteristics include: product categories, such as digital products, daily necessities and furniture, etc.; product prices, such as less than 100 yuan, 100 to 500 yuan, 500 to 1,000 yuan and more than 1,000 yuan; content quality, such as high quality, medium quality and low quality; content categories, such as articles, videos, pictures and texts, financial news, sports news, motorcycle travel guides, etc.; content keywords, such as "Kawasaki 400", "Beautiful Scenery of Tibet", etc.

[0066] As an improved solution, the recall algorithm generates a to-be-recommended list according to the product preference, the content preference and the content feature, further comprising:

[0067] The recall algorithm includes a real-time recall algorithm and a non-real-time recall algorithm;

[0068] The real-time recall algorithm calculates a first similarity between the content feature and the product preference and a second similarity between the content feature and the content preference through a collaborative filtering algorithm, and selects content to be recommended according to the first similarity and the second similarity to generate a first list to be recommended;

[0069] The non-real-time recall algorithm generates a second to-be-recommended list according to the product preference, the content preference and the content feature;

[0070] The to-be-recommended list is generated according to the first to-be-recommended list and the second to-be-recommended list.

[0071] As an improved solution, the real-time recall algorithm calculates the first similarity between the content feature and the product preference and the second similarity between the content feature and the content preference through a collaborative filtering algorithm, and selects the content to be recommended according to the first similarity and the second similarity to generate a first list to be recommended, further comprising:

[0072] The content features include product category and content category;

[0073] Calculating the first similarity between the commodity category and the commodity preference, and selecting content to be recommended according to the first similarity to generate a third list to be recommended;

[0074] calculating the second similarity between the content category and the content preference, and selecting content to be recommended according to the second similarity to generate a fourth list to be recommended;

[0075] The first list to be recommended is generated according to the third list to be recommended and the fourth list to be recommended.

[0076] In this embodiment, the collaborative filtering algorithms include the simrank algorithm and the word2vec algorithm; taking motorcycles as an example, the similarity evaluation criteria are as follows: if the product category is motorcycles and the product preference is Harley motorcycles, the first similarity is high; if the product category is motorcycles and the product preference is Apple mobile phones, the first similarity is low; if the content category is mobile phone reviews and the content preference is Apple mobile phone reviews, the second similarity is high; if the content category is mobile phone reviews and the content preference is book sharing, the second similarity is low; in order to ensure the real-time nature of user preferences, the third list to be recommended and the fourth list to be recommended only consider the product preferences and content preferences of the user within the past day, and the lengths of the third list to be recommended and the fourth list to be recommended are set according to the average value of the user's browsing volume.

[0077] As an improved solution, the non-real-time recall algorithm generates a second to-be-recommended list according to the product preference, the content preference and the content feature, further comprising:

[0078] The non-real-time recall algorithm includes a product recall algorithm and a content recall algorithm;

[0079] The content characteristics include content quality;

[0080] The product recall algorithm calculates a first recall score according to the product preference and the content quality weight; and selects content to be recommended according to the first recall score to generate a fifth list to be recommended;

[0081] The content recall algorithm generates a sixth to-be-recommended list according to the content preference and the content feature;

[0082] The second list to be recommended is generated according to the fifth list to be recommended and the sixth list to be recommended.

[0083] In this embodiment, the non-real-time recall algorithm is executed at least twice, one time is to execute the non-real-time recall algorithm on the user characteristics of the past two weeks, which is called the short-term recall algorithm; the other time is to execute the non-real-time recall algorithm on the user characteristics of the past three months, which is called the long-term recall algorithm.

[0084] As an improved solution, the content recall algorithm generates a sixth to-be-recommended list according to the content preference and the content feature, further comprising:

[0085] The content preference includes content category preference and content keyword preference;

[0086] Calculating a third recall score according to the content category preference and the content quality weight, and selecting to-be-recommended content according to the third recall score to generate a seventh to-be-recommended list;

[0087] Setting a word vector of a content keyword, calculating a semantic similarity between the content keyword preference and the word vector, and selecting content to be recommended according to the semantic similarity to generate an eighth list to be recommended;

[0088] The sixth list to be recommended is generated according to the seventh list to be recommended and the eighth list to be recommended.

[0089] In this embodiment, in order to ensure that the recalled content can reflect the diversity of user preferences, the lengths of the fifth to-be-recommended list to the eighth to-be-recommended list are set according to the maximum value of the user's browsing volume to meet the user's various needs.

[0090] In this embodiment, the content preference is reflected in the click rate of the content; there are problems in directly using the content keyword to select articles containing the word in the resource library, such as "event" and "match" are originally synonyms, but through the query, it is impossible to recall articles containing "match" through "event". So the word vector technology is used to solve this problem. If similar words are mapped to similar vectors, then we can recall semantically similar articles through semantic similarity calculation. However, in this process, the common word vectors in the industry, such as Baidu's ERNIE and Tencent's BERT, cannot meet the needs of the motorcycle vertical field; for example, we need the words "ninja400" (also known as "Kawasaki 400") and "CBR5005R" (also known as "Honda CBR500") to be close enough, and we also need the distance between "picking up the car" and "new car unboxing" to be close enough, so we need to train a word vector in the motorcycle field ourselves. This embodiment independently annotates the dictionary in the motorcycle field to ensure that these industry-specific words can be successfully segmented and extracted, and then uses the word vector training model (BERT) to train our corpus across the entire platform; ultimately, word vectors specific to the motorcycle field are generated, and the recalled articles can better reflect user preferences.

[0091] In step S300, samples are selected and the intent recognition model is trained based on the samples; the successfully trained intent recognition model derives the probability of the user's intent based on the user's characteristics.

[0092] As an improved solution, the sample selection step further includes:

[0093] The user characteristics include activity level and target behavior;

[0094] Setting an activity level threshold; selecting a target user, wherein the activity level of the target user exceeds the activity level threshold;

[0095] The samples include positive samples and negative samples. The target user who has the target behavior is taken as a positive sample, and the target user who does not have the target behavior is taken as a negative sample.

[0096] In this embodiment, the target behavior is an inquiry behavior; the activity level depends on the user's login frequency or the user's browsing volume or the user's posting volume within a period of time.

[0097] As an improved solution, the step of training the intention recognition model according to the sample further includes:

[0098] Set accuracy threshold;

[0099] The user features of the positive sample and the negative sample are passed through the intent recognition model to generate accuracy;

[0100] If the accuracy exceeds the accuracy threshold, the intention recognition model training is successful;

[0101] If the accuracy is lower than the accuracy threshold, the intent recognition model is modified and the next training is started.

[0102] In step S400, the sorting algorithm generates a recommendation list according to the probability and the list to be recommended.

[0103] In this embodiment, the sorting algorithm is the LightGBM algorithm, which is a framework for implementing the GBDT algorithm open sourced by Microsoft and supports efficient parallel training; GBDT is a long-standing model in machine learning, and its main idea is to use weak classifiers (decision trees) for iterative training to generate an optimal model, which has the advantages of good training effect and not easy to overfit. The sorting algorithm calculates the click-through rate based on the probability and the content features of the list to be recommended, sorts the content of the list to be recommended in descending order according to the click-through rate, and takes a preset number of contents with the highest probability to generate the recommendation list, and the preset number is determined according to the average number of user views.

[0104] In this embodiment, the content preference and the product preference are integrated through the recall algorithm to generate diverse preferences of users; and the target users are trained through the intent recognition model to generate probabilities of different user intentions, so as to facilitate content recommendation based on user intentions.

[0105] Example 2

[0106] This embodiment 2 provides a content recommendation system. Figure 3 , the content recommendation system includes: a configuration unit, an acquisition unit and an execution generation unit;

[0107] The configuration unit is used to configure the recall algorithm and the sorting algorithm; and to establish an intent recognition model;

[0108] The acquisition unit is used to acquire user characteristics and content characteristics;

[0109] The execution generation unit is used to execute the recall algorithm according to the user characteristics and the content characteristics to generate a list to be recommended; to run the intention recognition model according to the user characteristics to generate the probability of user intention; and to generate a recommendation list according to the probability and the list to be recommended.

[0110] As an improved solution, the configuration unit is used to select target users as samples according to the user characteristics, and train the intent recognition model according to the samples.

[0111] In this embodiment, the recommendation list is generated by the configuration unit and the execution unit according to the user characteristics and the content characteristics, so that content recommendation is more accurate.

[0112] Example 3

[0113] This embodiment 3 provides a computer-readable storage medium, which is used to store computer software instructions used to implement the content recommendation method described in the above embodiment 1, and includes a program for executing the above-mentioned content recommendation method; specifically, the executable program can be built into the content recommendation system 100, so that the content recommendation system 100 can implement the content recommendation method of the above embodiment 1 by executing the built-in executable program.

[0114] In addition, the computer-readable storage medium provided in this embodiment may adopt any combination of one or more readable storage media, wherein the readable storage medium includes electrical, optical, electromagnetic, infrared or semiconductor systems, systems or devices, or any combination of the above.

[0115] The serial numbers of the embodiments disclosed in the above embodiments are only for description and do not represent the advantages or disadvantages of the embodiments.

[0116] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made 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 content recommendation method, It is characterized in that The steps include: Configure the recall algorithm and sorting algorithm to establish the intent recognition model; Acquiring user characteristics and content characteristics, wherein the user characteristics include product preferences and content preferences; The recall algorithm generates a list to be recommended according to the product preference, the content preference and the content feature; Selecting samples, and training the intent recognition model according to the samples; The successfully trained intention recognition model derives the probability of the user's intention based on the user characteristics; The sorting algorithm generates a recommendation list according to the probability and the list to be recommended; The recall algorithm generates a to-be-recommended list according to the product preference, the content preference and the content feature, further comprising: The recall algorithm includes a real-time recall algorithm and a non-real-time recall algorithm; The real-time recall algorithm calculates a first similarity between the content feature and the product preference and a second similarity between the content feature and the content preference through a collaborative filtering algorithm, and selects content to be recommended according to the first similarity and the second similarity to generate a first list to be recommended; The non-real-time recall algorithm generates a second to-be-recommended list according to the product preference, the content preference and the content feature; generating the to-be-recommended list according to the first to-be-recommended list and the second to-be-recommended list; The real-time recall algorithm calculates the first similarity between the content feature and the product preference and the second similarity between the content feature and the content preference by a collaborative filtering algorithm, and selects the content to be recommended according to the first similarity and the second similarity to generate a first list to be recommended, further comprising: The content features include product category and content category; Calculating the first similarity between the commodity category and the commodity preference, and selecting content to be recommended according to the first similarity to generate a third list to be recommended; calculating the second similarity between the content category and the content preference, and selecting content to be recommended according to the second similarity to generate a fourth list to be recommended; generating the first list to be recommended according to the third list to be recommended and the fourth list to be recommended; The non-real-time recall algorithm generates a second to-be-recommended list according to the product preference, the content preference and the content feature, further comprising: The non-real-time recall algorithm includes a product recall algorithm and a content recall algorithm; The content characteristics include content quality; The product recall algorithm calculates a first recall score according to the product preference and the content quality weight; and selects content to be recommended according to the first recall score to generate a fifth list to be recommended; The content recall algorithm generates a sixth to-be-recommended list according to the content preference and the content feature; generating the second list to be recommended according to the fifth list to be recommended and the sixth list to be recommended; The step of generating a sixth to-be-recommended list according to the content preference and the content feature by the content recall algorithm further comprises: The content preference includes content category preference and content keyword preference; Calculating a third recall score according to the content category preference and the content quality weight, and selecting to-be-recommended content according to the third recall score to generate a seventh to-be-recommended list; Setting a word vector of a content keyword, calculating a semantic similarity between the content keyword preference and the word vector, and selecting content to be recommended according to the semantic similarity to generate an eighth list to be recommended; The sixth list to be recommended is generated according to the seventh list to be recommended and the eighth list to be recommended.

2. The content recommendation method according to claim 1, It is characterized in that The sample selection step further comprises: The user characteristics include activity level and target behavior; Setting an activity level threshold; selecting a target user, wherein the activity level of the target user exceeds the activity level threshold; The samples include positive samples and negative samples. The target user who has the target behavior is taken as a positive sample, and the target user who does not have the target behavior is taken as a negative sample.

3. The content recommendation method according to claim 2, It is characterized in that The step of training the intention recognition model according to the sample further includes: Set accuracy threshold; The user features of the positive sample and the negative sample are passed through the intent recognition model to generate accuracy; If the accuracy exceeds the accuracy threshold, the intention recognition model training is successful; If the accuracy is lower than the accuracy threshold, the intent recognition model is modified and the next training is started.

4. A content recommendation system, based on the content recommendation method according to claim 1, It is characterized in that The content recommendation system includes: a configuration unit, an acquisition unit and an execution generation unit; The configuration unit is used to configure the recall algorithm and the sorting algorithm; and to establish an intent recognition model; The acquisition unit is used to acquire user characteristics and content characteristics; The execution generation unit is used to execute the recall algorithm according to the user characteristics and the content characteristics to generate a list to be recommended; to run the intention recognition model according to the user characteristics to generate the probability of user intention; and to generate a recommendation list according to the probability and the list to be recommended.

5. The content recommendation system according to claim 4, It is characterized in that The configuration unit is used to select a target user according to the user characteristics, and train the intent recognition model according to the target user.

6. A computer storage medium, It is characterized in that Used to store computer software instructions used for the content recommendation method according to any one of claims 1 to 3, which includes a program designed for executing the content recommendation method.

Images