Big data AI e-commerce content recommendation method and management system

Abstract

The invention discloses a big data AI e-commerce content recommendation method and management system, and belongs to the technical field of e-commerce content recommendations, the system establishes an individual interest tag library of a user, and calculates the initial weight of each interest tag; and monitoring the real-time interest tag of the user clicking the content and the page opening time, and comparing with the interest tag library. On the premise that the real-time interest tag is matched with the interest tag library, if the page opening time exceeds a set threshold value, the weight of the interest tag is increased; otherwise, reducing the weight of the interest tag. And if the real-time interest tag is not matched with the interest tag library, adding the interest tag into the interest tag library, initializing the weight according to the page clicking time, and correspondingly reducing the weights occupied by other newly added interest tags. And randomly selecting and combining interest labels according to the weights, and recommending commodities containing the combined interest labels. The method can monitor the interest change of the user in real time, so that the commodity recommendation is more accurate.

Description

Technical Field

[0001] The present invention belongs to the technical field of e-commerce content recommendation, and specifically, relates to a big data AI e-commerce content recommendation method and management system. Background Art

[0002] With the popularization of the Internet and the rapid development of e-commerce, big data and artificial intelligence technologies have begun to be widely used in the field of e-commerce. This method is based on the user's historical behavior data, purchase records, preferences and other information, and uses complex algorithm models to mine and analyze deep information to achieve accurate control of user personalized recommendations.

[0003] When recommending products, the e-commerce recommendation system in the prior art often pays more attention to the user's historical information, including historical purchase records, historical browsing records, historical preferences, etc., and recommends some products of this type or related to this type to the user based on this information. This method ignores the fact that people's interests are dynamically changing, and they will not always be interested in certain products. The system's constant recommendation of products of this type will cause user selection fatigue, resulting in poor final recommendation results. In order to solve the above problem, a solution is provided that can adjust the recommended products according to the real-time changes in user interests, effectively improving the accuracy of recommendations. The present invention provides the following technical solutions. Summary of the invention

[0004] The purpose of the present invention is to provide a big data AI e-commerce content recommendation method and management system to solve the problem that the e-commerce recommendation system in the prior art spends a lot of time paying attention to the user's historical information to make product recommendations when recommending products to users, which leads to user selection fatigue, and cannot adjust the recommended products according to the real-time changes in user interests.

[0005] The purpose of the present invention can be achieved through the following technical solutions:

[0006] The big data AI e-commerce content recommendation method is characterized in that the method specifically includes the following steps:

[0007] S1. Establish a user's interest tag library and calculate the initial weight of each interest tag in the interest tag library;

[0008] S2. Monitor the real-time interest tags and page click time of the user's clicked content, eliminate invalid page click time, and determine whether the real-time interest tags match the interest tag library. If they match, the click behavior is called a positive click; otherwise, it is called a reverse click;

[0009] S3. For positive clicks, the weight of the interest tag is adjusted according to the length of time the page is clicked;

[0010] S4. For reverse clicks, add the interest tag to the interest tag library, initialize the weight of the interest tag according to the page click time, and reduce the weights of other newly added interest tags accordingly;

[0011] S5. Randomly select and combine interest tags according to weights, and recommend products containing the combined interest tags to users.

[0012] As a further solution of the present invention, the method for calculating the initial weight of each interest tag in the interest tag library is:

[0013] S11, obtain the corpus and select a specific document D;

[0014] S12, classify the interest tags of document D, and obtain a list of all interest tags in document D;

[0015] S13, calculate the frequency TF(t,D) of each interest tag appearing in document D;

[0016] S14, counting the number of documents df(t) in which each interest tag appears in the corpus, calculating the total number N of all documents in the corpus, and calculating the discrimination ability of each interest tag t in the entire corpus according to the formula IDF(t)=log(N / (1+df(t)));

[0017] S15. For each interest tag in the document D, calculate its initial weight according to the formula w(t)=TF(t,D)*IDF(t).

[0018] As a further solution of the present invention, the selected specific document D is a product evaluation.

[0019] As a further solution of the present invention, a method for eliminating invalid page opening time is:

[0020] Collect the user's historical purchase and collection information within the recent period T, including the number of times the user scrolled the screen by hand when browsing each product page g gun , duration of each scroll gun and distance s gun , the number of times a link or image is clicked during browsing zhon and time t zhon ;

[0021] According to the formula The average scrolling speed of each product in each period of time is obtained, and this average is recorded as the user's usual browsing speed. The abnormal interval of browsing speed within time T is calculated by the interquartile range method, and the abnormal values ​​in the browsing speed are eliminated. The interval of the user's usual browsing speed is obtained as [v min ,v max ];

[0022] According to the formula Obtain the total duration of the user clicking on the link or image, record this total duration as the user's usual in-depth browsing duration, calculate the abnormal interval of the in-depth browsing duration within time T by using the interquartile range method, remove the abnormal values ​​in the in-depth browsing duration, and obtain the interval of the user's usual in-depth browsing duration [t min ,t max ];

[0023] Get the user's browsing speed v for the current product now , the in-depth browsing time is t sustain ;

[0024] like or The time when the page is clicked is removed, and the weights of the interest tags in the interest tag library remain unchanged.

[0025] As a further solution of the present invention, for positive clicks, the method for adjusting the weight of interest tags according to the length of time the page is clicked is:

[0026] Set the user's interest tag library Ij (j = 1, 2, ..., m), and the corresponding initialization weight is Wi (i = 1, 2, ..., m);

[0027] Obtain the user page click time Δt=t2-t1, and adjust the weight of the interest tag according to the formula Wi'=Wi(1+(-1)^E*0.01*Δt);

[0028] Where t1 represents the time when the user starts to click on the page, t2 represents the time when the user closes the page, Wi' represents the adjusted interest tag weight, Wi represents the unadjusted interest tag weight, and E represents a piecewise function;

[0029] If Δt>η, it indicates that the user is interested in this type of product. The weight of the interest tag is increased and the value of E is set to 2. If Δt<=η, it indicates that the user is not interested in this type of product, and the value of E is set to 1.

[0030] As a further embodiment of the present invention, the value of η is 1.

[0031] As a further solution of the present invention, for reverse clicks, the interest tag needs to be added to the interest tag library, and the calculation method for adjusting the initial weight of the added interest tag according to the page click time is:

[0032] Add a new label Lj, according to the formula Wj'=γ*Δt, γ represents the weight adjustment factor, which is adjusted by the setter.

[0033] As a further solution of the present invention, when a user performs a reverse click, the weight of other newly added interest tags needs to be reduced according to the time decay situation. The specific method is:

[0034] According to Δt'=tnow-torigin, the difference between the newly added interest tag and the current time is obtained;

[0035] Among them, tnow represents the current time, and torigin represents the time when the interest tag library is added;

[0036] Since the latest added interest tags can better reflect the changes in users' real-time interests, the interest tags added previously will lose the attention of users as time goes by. According to the formula Wt''=((e^-Δt') / Δt')*Wt', the weight of the interest tags added previously changes over time.

[0037] Among them, Wt' represents the weight of the interest tags added previously, and Wt" represents the weight of the interest tags after the change over time;

[0038] As a further solution of the present invention, the method of randomly selecting and combining interest tags according to weights is:

[0039] According to the formula Calculate the probability of selecting each interest tag;

[0040] in, Represents the sum of all interest tag weights, and the normalized weight Pi satisfies

[0041] According to the formula Construct a discrete distribution function to convert the normalized weights into intervals;

[0042] Among them, F(ti) represents the cumulative probability from the first interest tag to the i-th interest tag;

[0043] Generate a random number r between 0 and 1, and find the first cumulative distribution function value greater than or equal to r. This is the selected interest tag.

[0044] The selected interest tags are combined, and the products containing the group of interest tags are found for recommendation.

[0045] A big data AI e-commerce content recommendation system, comprising: an e-commerce tag collection unit, an e-commerce tag comparison unit, an e-commerce tag combination unit, and an e-commerce content recommendation unit;

[0046] An e-commerce tag collection unit, which collects the user's historical purchase data and purchase preferences to establish an interest tag library, and transmits the interest tag library to the e-commerce tag comparison unit;

[0047] An e-commerce tag comparison unit, which monitors the user's real-time interest tags and page click time, removes invalid page click time, and if the real-time interest tag matches the interest tag library, dynamically adjusts the weight of the interest tag according to the page click time. Otherwise, the real-time interest tag is added to the interest tag library, and the interest tag weight is initialized according to the page click time, and the weight of the real-time interest tag added previously is reduced accordingly, and the weight of each interest tag finally obtained is transmitted to the e-commerce tag combination unit;

[0048] An e-commerce tag combination unit, which is used to randomly select and combine interest tags according to weights, and transmit the combined interest tags to the e-commerce recommendation unit;

[0049] The e-commerce content recommendation unit is used to match the combined interest tags with suitable products and recommend the products to users in real time.

[0050] Beneficial effects of the present invention:

[0051] 1. The present invention takes into account that user interests change dynamically, and always recommending products based on user historical information will cause selection fatigue to users, resulting in the problem of poor final recommendation effect. Based on the real-time clicks of users on products, the interest tags of the products and the page opening time are fully explored, and the interest tags are compared with the interest tag library. If they match, the weight of the interest tags is adjusted according to the length of the page opening time. If they do not match, the interest tags are added to the interest tag library, and their weight is determined according to the page opening time, and the weight of the newly added interest tags is reduced accordingly. Finally, interest tags are randomly selected and combined according to the weights, and products containing the combined interest tags are recommended to users, thereby improving the real-time and accuracy of product recommendations.

[0052] 2. The present invention fully considers the influence of time factors on newly added interest tags. As time goes by, the influences of previously added interest tags and currently added interest tags on users are different. The closer to the user's click time, the more it can reflect the user's current interest status. By introducing an exponential function to calculate the weights of previously added interest tags, the weights of interest tags that are more in line with the actual situation are obtained, which prepares for the subsequent combination of interest tags according to the weights.

[0053] 3. The present invention randomly selects and combines interest tags according to weights, which can provide users with diverse and potential products of interest, thereby improving users' satisfaction with the system's recommended content. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a schematic diagram of the process of the present invention;

[0055] Figure 2 It is a system block diagram of the present invention. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] like Figure 1 , big data AI e-commerce content recommendation method, including:

[0058] S1. Collect and summarize the user's interest tag library and calculate the initial weight of each interest tag in the interest tag library;

[0059] The user's interest tag library mainly considers product quality, brand reputation, function and performance, logistics speed, etc.

[0060] Product quality is divided into four categories: low, medium, good, and excellent, corresponding to 1 to 4 stars respectively, and is mainly determined based on the experience scores of each user after actual purchase;

[0061] Brand reputation is divided into three categories: low, medium and high. It is affected by five dimensions: brand history and popularity, product quality and service level, consumer satisfaction and loyalty, social responsibility and brand image, and advertising and marketing promotion. The weights are set to 20%, 30%, 25%, 15% and 10% respectively. The above five dimensions can be scored between 0 and 100, and then the brand reputation score is calculated = (brand history and popularity score × 20%) + (product quality and service level score × 30%) + (consumer satisfaction and loyalty score × 25%) + (social responsibility and brand image score × 15%) + (advertising and marketing promotion score × 10%). The score below 60 is low, between 60 and 89 is medium, and between 90 and 100 is high.

[0062] Functions and performance are divided into three categories: extremely poor, poor, good, and excellent. They are affected by the four dimensions of core function realization, performance, usability and user experience, and compatibility and scalability. The weights are set to 30%, 40%, 20%, and 10% respectively. The above four dimensions can be scored between 0 and 100, and the function and performance score is calculated as (core function realization score × 30%) + (performance score × 40%) + (usability and user experience score × 20%) + (compatibility and scalability score × 10%). The score below 60 is extremely poor, 60 to 75 is good, 76 to 89 is good, and 90 to 100 is excellent.

[0063] The logistics speed is divided into three types: slow, normal, and fast, corresponding to 1 to 3 stars respectively. Users can rate the time taken from ordering to receiving the goods;

[0064] The method for calculating the initial weight of each interest tag in the interest tag library is:

[0065] S11, obtain the corpus and select a specific document D;

[0066] The corpus is a large text database that contains a large number of documents. These documents can come from various sources, but they are all related to the product. For example, in the product evaluation area, it contains all the information we need, such as product quality, function and performance, logistics speed, etc., which can show our attention to certain aspects of the product. In the product introduction area, we can also obtain relevant information such as the brand reputation and popularity of the product, which can be used as the basis for the initial weight calculation of the product interest tag later. The specific document is a document selected from the corpus that can represent the user's purchase situation and the next purchase intention. It plays a vital role in adjusting the product content recommendation according to the user's actual interest.

[0067] In one embodiment of the present invention, a product review is selected as a specific document D;

[0068] S12, classify the interest tags of document D, and obtain a list of all interest tags in document D;

[0069] S13, counting the number of times each interest tag appears in document D count(t), and calculating the total number of words in document D |D|, and calculating the frequency of each interest tag appearing in document D according to the formula TF(t,D)=count(t) / |D|;

[0070] The method to determine the total number of words in a Chinese document is:

[0071] If the document contains characters that do not need to be calculated, such as punctuation marks, numbers, special symbols, etc., you can perform preprocessing to remove these characters;

[0072] Since there are no obvious spaces between Chinese words, word segmentation is required. This can be done by using Chinese word segmentation tools or software. Word segmentation tools can split the text into independent words, which is convenient for subsequent statistics.

[0073] After the word segmentation is completed, the total number of words in the document is counted. This is usually achieved by traversing the word segmentation results and counting the number of different words in them;

[0074] S14, counting the number of documents df(t) in which each interest tag appears in the corpus, calculating the total number N of all documents in the corpus, and calculating the discrimination ability of each interest tag t in the entire corpus according to the formula IDF(t)=log(N / (1+df(t)));

[0075] When the IDF is calculated, 1 is added to the denominator to avoid taking 0, which enhances the robustness of the algorithm. The log function is added to prevent problems such as weight explosion, which provides mathematical intuition and convenience.

[0076] S15. For each interest tag in the document D, calculate its initial weight according to the formula w(t)=TF(t,D)*IDF(t);

[0077] S2. Monitor the real-time interest tags and page click time of the user's clicked content, eliminate invalid page click time, and determine whether the real-time interest tags match the interest tag library. If they match, the click behavior is called a positive click; otherwise, it is called a reverse click;

[0078] The real-time interest tags of the content clicked by the user are monitored. The real-time interest tags are extracted from the product page clicked by the user, and the interest tags displayed in front of the product page are focused on. These interest tags can better reflect the user's attention to it;

[0079] The page click time refers to the effective time when a user clicks on a product page to browse product information. If the product page is stuck or crashes, it cannot accurately represent the time when the user clicks on the product page.

[0080] The method to eliminate invalid page click time is:

[0081] Collect the user's historical purchase and collection information within the recent period T, including the number of times the user scrolled the screen by hand when browsing each product page g gun , duration of each scroll gun and distance s gun , the number of times a link or image is clicked during browsing zhon and time t zhon ;

[0082] The number and duration of clicks on links or images during browsing can be calculated using the browser's PAI. Before clicking a link, call the window.performance.now() method to get a timestamp as the start time. When entering a new link page, call window.performance.now() again at an appropriate time to get the end timestamp, and then calculate the duration t. zhon, by calculating half of the total number of calls to window.performance.now() as the number g zhon ;

[0083] According to the formula The average scrolling speed of each product in each period of time is obtained, and this average is recorded as the user's usual browsing speed. The abnormal interval of browsing speed within time T is calculated by the interquartile range method, and the abnormal values ​​in the browsing speed are eliminated. The interval of the user's usual browsing speed is obtained as [v min ,v max ];

[0084] According to the formula Obtain the total duration of the user clicking on the link or image, record this total duration as the user's usual in-depth browsing duration, calculate the abnormal interval of the in-depth browsing duration within time T by using the interquartile range method, remove the abnormal values ​​in the in-depth browsing duration, and obtain the interval of the user's usual in-depth browsing duration [t min ,t max ];

[0085] The interquartile range method is used to calculate the abnormal interval and eliminate abnormal values. For example, a set of data is 12, 5, 9, 18, 3, 21, 7. First, sort them to get: 3, 5, 7, 9, 12, 18, 21. Since the number of this set of data is an odd number, the first quartile Q1 = 5.5 and the third quartile Q3 = 20.25 are calculated by linear interpolation. The interquartile range IQR = Q3-Q1 = 14.75 is calculated to determine the limits of the abnormal values. The lower limit = Q1-1.5*IQR = -16.625, the upper limit = Q3+1.5*IQR = 42.375, and the values ​​less than the lower limit and greater than the upper limit are classified as abnormal values ​​and eliminated;

[0086] Get the user's browsing speed v for the current product now , the in-depth browsing time is t sustain ;

[0087] like or The page click time is removed, and the weights of each interest tag in the interest tag library remain unchanged;

[0088] S3. For positive clicks, the weight of the interest tag is adjusted according to the length of time the page is clicked;

[0089] The method for dynamically adjusting the weight of interest tags is:

[0090] Set the user's interest tag library Ij (j = 1, 2, ..., m), and the corresponding initialization weight is Wi (i = 1, 2, ..., m);

[0091] Obtain the user page click time Δt=t2-t1, and adjust the weight of the interest tag according to the formula Wi'=Wi(1+(-1)^E*0.01*Δt);

[0092] Where t1 represents the time when the user starts to click on the page, t2 represents the time when the user closes the page, Wi' represents the adjusted interest tag weight, Wi represents the unadjusted interest tag weight, and E represents a piecewise function;

[0093] If Δt>η, it means that the user is interested in this type of product, and the weight of the interest tag is increased, and the value of E is set to 2. If Δt<=η, it means that the user is not interested in this type of product, and the value of E is set to 1;

[0094] η is the threshold value set to determine whether a user is interested in a product. If it is set too small, it is easy to classify products that the user is not interested in as being of interest; if it is set too large, it is easy to classify products that the user is interested in as being of no interest. Both situations will result in poor recommendation results and affect the user's final choice.

[0095] In one embodiment of the present invention, the value of n is 1;

[0096] The present invention takes into account that user interests change dynamically. Recommending products based on user history information for a long time will cause user selection fatigue, resulting in poor recommendation results. By using the user's real-time click content and the length of time the page is opened, the dynamic changes of user interests can be tracked in real time to adjust the recommended content, thereby improving the accuracy of recommendations to users.

[0097] S4. For reverse clicks, add the interest tag to the interest tag library, initialize the weight of the interest tag according to the page click time, and reduce the weights of other newly added interest tags accordingly;

[0098] The calculation method for dynamically adjusting the initial weight of the added interest tags is:

[0099] Add a new interest tag Lj, according to the formula Wj'=γ*Δt, γ represents the weight adjustment factor;

[0100] In one embodiment of the present invention, γ is 0.01;

[0101] When a user clicks back, the weight of other newly added interest tags needs to be reduced according to the time decay. The specific method is:

[0102] According to Δt'=tnow-torigin, the difference between the newly added interest tag and the current time is obtained;

[0103] Among them, tnow represents the current time, and torigin represents the time when the interest tag library is added;

[0104] Since the latest added interest tags can better reflect the changes in users' real-time interests, the interest tags added previously will lose the attention of users as time goes by. According to the formula Wt''=((e^-Δt') / Δt')*Wt', the weight of the interest tags added previously changes over time.

[0105] Among them, Wt' represents the weight of the interest tags added previously, and Wt" represents the weight of the interest tags after the change over time;

[0106] The present invention fully considers the influence of time factors on newly added interest tags. As time goes by, the influence of previously added interest tags and currently added interest tags on users is different. The closer the time of user click, the more it can reflect the user's current interest state. By introducing an exponential function to calculate the weight of previously added interest tags, the weight of interest tags that is more in line with the actual situation is obtained.

[0107] S5. Randomly select and combine interest tags according to weights, and recommend products containing the combined interest tags to the user;

[0108] The method of randomly selecting and combining interest tags by weight is:

[0109] According to the formula Calculate the probability of selecting each interest tag;

[0110] in, Represents the sum of all interest tag weights, and the normalized weight Pi satisfies

[0111] According to the formula Construct a discrete distribution function to convert the normalized weights into intervals;

[0112] Among them, F(ti) represents the cumulative probability from the first interest tag to the i-th interest tag;

[0113] Generate a random number r between 0 and 1, and find the first cumulative distribution function value greater than or equal to r. This is the selected interest tag.

[0114] Combine the selected interest tags and find out the products containing the group of interest tags for recommendation;

[0115] like Figure 2 , a big data AI e-commerce content recommendation system, comprising: an e-commerce tag collection unit, an e-commerce tag comparison unit, an e-commerce tag combination unit, and an e-commerce content recommendation unit;

[0116] An e-commerce tag collection unit, which collects the user's historical purchase data and purchase preferences to establish an interest tag library, and transmits the interest tag library to the e-commerce tag comparison unit;

[0117] An e-commerce tag comparison unit, which monitors the user's real-time interest tags and page click time, removes invalid page click time, and if the real-time interest tag matches the interest tag library, dynamically adjusts the weight of the interest tag according to the page click time. Otherwise, the real-time interest tag is added to the interest tag library, and the interest tag weight is initialized according to the page click time, and the weight of the real-time interest tag added previously is reduced accordingly, and the weight of each interest tag finally obtained is transmitted to the e-commerce tag combination unit;

[0118] An e-commerce tag combination unit, which is used to randomly select and combine interest tags according to weights, and transmit the combined interest tags to the e-commerce recommendation unit;

[0119] The e-commerce content recommendation unit is used to match the combined interest tags with suitable products and recommend the products to users in real time.

[0120] In the description of the specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0121] The above contents are merely examples and explanations of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the invention or exceed the scope defined by the claims, they shall all fall within the protection scope of the present invention.

Claims

1. A big data AI e-commerce content recommendation method, characterized in that: The method specifically comprises the following steps: S1. Establish a user's interest tag library and calculate the initial weight of each interest tag in the interest tag library; S2. Monitor the real-time interest tags and page click time of the user's clicked content, eliminate invalid page click time, and determine whether the real-time interest tags match the interest tag library. If they match, the click behavior is called a positive click; otherwise, it is called a reverse click; S3. For positive clicks, the weight of the interest tag is adjusted according to the length of time the page is clicked; S4. For reverse clicks, add the interest tag to the interest tag library, initialize the weight of the interest tag according to the page click time, and reduce the weights of other newly added interest tags accordingly; S5. Randomly select and combine interest tags according to weights, and recommend products containing the combined interest tags to users.

2. The big data AI e-commerce content recommendation method according to claim 1 is characterized in that: The method for calculating the initial weight of each interest tag in the interest tag library is: S11, obtain the corpus and select a specific document D; S12, classify the interest tags of document D, and obtain a list of all interest tags in document D; S13, calculate the frequency TF(t,D) of each interest tag appearing in document D; S14, counting the number of documents df(t) in which each interest tag appears in the corpus, calculating the total number N of all documents in the corpus, and calculating the discrimination ability of each interest tag t in the entire corpus according to the formula IDF(t)=log(N / (1+df(t))); S15. For each interest tag in the document D, calculate its initial weight according to the formula w(t)=TF(t,D)*IDF(t).

3. The big data AI e-commerce content recommendation method according to claim 2 is characterized in that: The selected specific document D is a product review.

4. The big data AI e-commerce content recommendation method according to claim 1, characterized in that: The method to eliminate invalid page click time is: Collect the user's historical purchase and collection information within the recent period T, including the number of times the user scrolled the screen by hand when browsing each product page g gun , duration of each scroll gun and distance s gun , the number of times a link or image is clicked during browsing zhon and time t zhon ; According to the formula The average scrolling speed of each product in each period of time is obtained, and this average is recorded as the user's usual browsing speed. The abnormal interval of browsing speed within time T is calculated by the interquartile range method, and the abnormal values ​​in the browsing speed are eliminated. The interval of the user's usual browsing speed is obtained as [v min ,v max ]; According to the formula Obtain the total duration of the user clicking on the link or image, record this total duration as the user's usual in-depth browsing duration, calculate the abnormal interval of the in-depth browsing duration within time T by using the interquartile range method, remove the abnormal values ​​in the in-depth browsing duration, and obtain the interval of the user's usual in-depth browsing duration [t min ,t max ]; Get the user's browsing speed v for the current product now , the in-depth browsing time is t sustain ; like or The time when the page is clicked is removed, and the weights of the interest tags in the interest tag library remain unchanged.

5. The big data AI e-commerce content recommendation method according to claim 1, characterized in that: For positive clicks, the method to adjust the weight of interest tags according to the length of time the page is clicked is: Set the user's interest tag library Ij (j = 1, 2, ..., m), and the corresponding initialization weight is Wi (i = 1, 2, ..., m); Obtain the user page click time Δt=t2-t1, and adjust the weight of the interest tag according to the formula Wi'=Wi(1+(-1)^E*0.01*Δt); Where t1 represents the time when the user starts to click on the page, t2 represents the time when the user closes the page, Wi' represents the adjusted interest tag weight, Wi represents the unadjusted interest tag weight, and E represents a piecewise function; If Δt>η, it indicates that the user is interested in this type of product. The weight of the interest tag is increased and the value of E is set to 2. If Δt<=η, it indicates that the user is not interested in this type of product, and the value of E is set to 1.

6. The big data AI e-commerce content recommendation method according to claim 5, characterized in that: The value of η is 1.

7. The big data AI e-commerce content recommendation method according to claim 1, characterized in that: For reverse clicks, the interest tag needs to be added to the interest tag library. The calculation method for adjusting the initial weight of the added interest tag according to the page click time is: Add a new label Lj, according to the formula Wj'=γ*Δt, γ represents the weight adjustment factor, which is adjusted by the setter.

8. The big data AI e-commerce content recommendation method according to claim 1, characterized in that: When a user clicks back, the weight of other newly added interest tags needs to be reduced according to the time decay. The specific method is: According to Δt'=tnow-torigin, the difference between the newly added interest tag and the current time is obtained; Among them, tnow represents the current time, and torigin represents the time when the interest tag library is added; According to the formula Wt” = ((e^-Δt') / Δt')*Wt', find out how the weight of the interest tags added previously changes over time; Among them, Wt' represents the weight of the interest tags added previously, and Wt" represents the weight of the interest tags after the change over time.

9. The big data AI e-commerce content recommendation method according to claim 1, characterized in that: The method of randomly selecting and combining interest tags by weight is: According to the formula Calculate the probability of selecting each interest tag; in, Represents the sum of all interest tag weights, and the normalized weight Pi satisfies According to the formula Construct a discrete distribution function to convert the normalized weights into intervals; Among them, F(ti) represents the cumulative probability from the first interest tag to the i-th interest tag; Generate a random number r between 0 and 1, and find the first cumulative distribution function value. The interest tag with a value greater than or equal to r is the selected interest tag. The selected interest tags are combined and the products containing the group of interest tags are found for recommendation.

10. A big data AI e-commerce content recommendation system, used to execute the big data AI e-commerce content recommendation method according to any one of claims 1 to 9, characterized in that: The recommendation system includes: an e-commerce tag collection unit, an e-commerce tag comparison unit, an e-commerce tag combination unit, and an e-commerce content recommendation unit; An e-commerce tag collection unit, which collects the user's historical purchase data and purchase preferences to establish an interest tag library, and transmits the interest tag library to the e-commerce tag comparison unit; An e-commerce tag comparison unit, which monitors the user's real-time interest tags and page click time, removes invalid page click time, and if the real-time interest tag matches the interest tag library, dynamically adjusts the weight of the interest tag according to the page click time. Otherwise, the real-time interest tag is added to the interest tag library, and the interest tag weight is initialized according to the page click time, and the weight of the real-time interest tag added previously is reduced accordingly, and the weight of each interest tag finally obtained is transmitted to the e-commerce tag combination unit; An e-commerce tag combination unit, which is used to randomly select and combine interest tags according to weights, and transmit the combined interest tags to the e-commerce recommendation unit; The e-commerce content recommendation unit is used to match the combined interest tags with suitable products and recommend the products to users in real time.

Images