A news content recommendation method based on multidimensional feature analysis

Through multi-dimensional feature analysis and user portrait adjustment, the lack of diversity and cold start problems in news content recommendations are solved, and the balance between personalization and novelty is achieved, and the user experience is improved.

CN118626716BActive Publication Date: 2024-11-08NANJING FORESTRY UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410723268.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-05
Publication Date
2024-11-08
Estimated Expiration
2044-06-05

AI Technical Summary

Technical Problem

There are insufficient diversity, cold start and sparseness in the existing news content recommendation methods. It is difficult for fixed models to adapt to changes in user interests and it is difficult to balance personalization and novelty.

Method used

The news content recommendation method based on multi-dimensional feature analysis is adopted, and the appropriate recommendation model is selected through user classification and category marking, and the model is corrected based on the access information to form a target user portrait, and the recommended content is adjusted in real time to adapt to changes in user interests.

Benefits of technology

It effectively avoids the problems of insufficient diversity and cold start, improves the rationality and pertinence of content recommendations in new fields, ensures a balance between personalization and novelty, and improves user experience and satisfaction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118626716B_ABST
    Figure CN118626716B_ABST
Patent Text Reader

Abstract

The present invention belongs to the technical field of news content recommendation, and specifically discloses a news content recommendation method based on multidimensional feature analysis, which includes: collecting registration information and access information of target users; classifying target users and marking categories; selecting a benchmark recommendation model that matches the target user from a preset benchmark recommendation model library; correcting the benchmark recommendation model; building a target user portrait; analyzing the interest dynamics of the target user, and setting the monitoring adjustment frequency of the target user portrait; uploading the target user portrait, the monitoring adjustment frequency of the target user portrait, and the target recommendation model to a news push platform. The present invention avoids various problems existing in the current two fixed recommendation models, ensures the rationality of the new field content recommendation settings and the pertinence of the recommended content, and at the same time, by setting the monitoring adjustment frequency, it is convenient to adjust the recommended content in a timely manner to maintain consistency with user interests.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of news content recommendation, and relates to a news content recommendation method based on multidimensional feature analysis. Background Art

[0002] News is a form of media that delivers the latest information, reports social events, and spreads opinions. In the era of information explosion, the accuracy, timeliness, and objectivity of news are increasingly required, and news media has become one of the main channels for people to understand the world. In this context, the importance of news content recommendation processing is highlighted.

[0003] Currently, news content recommendation is mainly divided into two main types: content-based recommendation and collaborative filtering recommendation. Content-based recommendation system analyzes the text content, tags and other features of the article, and then matches similar content according to the user's interests for recommendation. Collaborative filtering recommendation predicts the content that the user may be interested in by analyzing the user's behavior data, such as clicks, readings, comments, etc.

[0004] There are still several deficiencies in the two current news content recommendation methods: 1. Recommendations based solely on content will lead to insufficient diversity in the recommended content. The current tendency is to continuously reinforce users' existing interests and lack recommendations for content in new areas, which makes users always in an "information cocoon" and makes it difficult for them to come into contact with new areas of interest.

[0005] 2. Recommendations based solely on collaborative filtering have the cold start problem, i.e., there is not enough data to recommend new users or new news, and the sparsity problem, i.e., the interactions between users and news are very rare, making it difficult to find similar users or news.

[0006] 3. Comprehensive content and collaborative filtering recommendations will make the hybrid system more complex, and it is a challenge to balance and integrate different recommendation algorithms.

[0007] 4. The current fixed model recommendation method is not only difficult to quickly adapt to changes in user interests, but also unable to improve the pertinence of recommended content, and also unable to ensure the balance between personalization and novelty of recommended content. Summary of the invention

[0008] In view of this, in order to solve the problems raised in the above background technology, a news content recommendation method based on multidimensional feature analysis is proposed.

[0009] The purpose of the present invention can be achieved through the following technical solutions: The present invention provides a news content recommendation method based on multidimensional feature analysis, the method comprising: collecting registration information and access information of target users.

[0010] According to the registration information and access information of the target user, the target user is classified and tagged.

[0011] According to the category tag of the target user, a benchmark recommendation model matching the target user is selected from a preset benchmark recommendation model library.

[0012] The benchmark recommendation model is modified according to the access information of the target user, and the modified benchmark recommendation model is used as the target recommendation model for the target user.

[0013] Personal information is extracted from the target user's registration information, and the personal information, category label and target recommendation model are associated to form a target user profile.

[0014] Analyze the interest trends of target users and set the monitoring and adjustment frequency of target user portraits.

[0015] Upload the target user portrait, the monitoring adjustment frequency of the target user portrait, and the target recommendation model to the news push platform.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention avoids the multi-faceted deficiencies, cold start and sparsity problems of the two current fixed recommendation models by performing user classification and category labeling, and selecting and modifying the recommendation model based on the user's category labeling, thereby ensuring the rationality of the new field content recommendation setting, reducing the difficulty for users to access new areas of interest, and also avoiding the integration challenges and trade-off challenges of the comprehensive recommendation of the two models.

[0017] (2) The present invention determines the adjustment weight by analyzing the browsing situation of news from different sources, and confirms the adjustment value of the adjustment weight, and then corrects the recommendation model, avoiding the shortcomings of the current fixed model recommendation method. It can not only adapt to the changes in user interests in a timely manner, but also ensure the pertinence of the recommended content as much as possible, thereby effectively maintaining the balance between the personalization and novelty of the recommended content.

[0018] (3) The present invention further classifies and judges when the registration time is in the interval between the registration time of two categories, thereby effectively making up for the deficiency that the current classification based only on the registration time may not accurately distinguish users, and realizes the in-depth classification of users, improves the accuracy and adaptability of user classification, and also ensures the pertinence and effectiveness of subsequent user news push, thereby facilitating the improvement of user reading experience.

[0019] (4) The present invention analyzes the access continuity and participation effectiveness based on time, and then confirms the activity level and tendency category analysis, which solves the shortcomings of traditional activity analysis that only analyzes based on user login frequency and simple behavioral indicators. It fully reflects the actual participation of users on the platform, and can more accurately evaluate the user's activity level, thereby realizing personalized activity level classification and accurate user tendency category analysis.

[0020] (5) The present invention analyzes the source change index value and the topic change index value, and then sets the monitoring adjustment frequency of the target user portrait, so as to capture the changes in the user's preference for news sources and topics, so as to adjust the recommended content in time and maintain consistency with the user's interests, thereby improving the accuracy and precision of the recommendation model, and at the same time being able to provide personalized recommendation content that better meets user needs, thereby improving user experience and satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0022] Figure 1 The present invention is a schematic flow chart of the steps for implementing the method. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. 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.

[0024] See also Figure 1 As shown, the present invention provides a news content recommendation method based on multi-dimensional feature analysis, the method comprising: S1, collecting registration information and access information of target users.

[0025] S2. Classify the target user according to the registration information and access information of the target user, and mark the category.

[0026] S3. According to the category label of the target user, a benchmark recommendation model matching the target user is selected from a preset benchmark recommendation model library.

[0027] It should be added that the benchmark recommendation model library stores benchmark recommendation models of categories, where the benchmark recommendation model models are derived through big data analysis and user behavior research, such as content-based recommendation models and collaborative filtering recommendation models.

[0028] S4. According to the access information of the target user, the benchmark recommendation model is modified, and the modified benchmark recommendation model is used as the target recommendation model for the target user.

[0029] The embodiments of the present invention classify users and mark categories, and select and modify recommendation models based on user category labels, thereby avoiding the multi-faceted deficiencies, cold start and sparsity problems of the two current fixed recommendation models, ensuring the rationality of the new field content recommendation settings, reducing the difficulty for users to access new areas of interest, and also avoiding the integration challenges and trade-off challenges of the comprehensive recommendation of the two models.

[0030] S5. Extract personal information from the target user's registration information, associate the personal information, category label and target recommendation model, and form a target user profile.

[0031] S6. Analyze the interest dynamics of target users and set the monitoring and adjustment frequency of target user portraits.

[0032] S7. Upload the target user portrait, the monitoring adjustment frequency of the target user portrait, and the target recommendation model to the news push platform.

[0033] Specifically, the specific process of step S2 is as follows: classify the target users, including: S2-1, based on the pre-set registration time intervals corresponding to each category, sort each category from small to large according to its registration time, and record them as category A, category B and category C in turn.

[0034] In a specific embodiment, each category in a specific actual scenario may be new users, active users and old users. Specifically, the registration time interval of new users is within one month, the actual registration time interval of old users is more than six months, and the registration time interval of active users is 1 month to 6 months, that is, category A corresponds to new users, category B corresponds to active users, and category C corresponds to old users.

[0035] S2-2. Extract the registration duration from the registration information of the target user.

[0036] S2-3. If the registration duration of the target user is within the registration duration interval of a certain category, the category is used as the category of the target user.

[0037] S2-4. If the registration duration of the target user is at the intersection of the registration duration interval of a certain category and the registration duration interval of another category, the two categories are combined into a judgment category group to judge the category tendency of the target user, and then the tendency category is used as the category of the target user.

[0038] The embodiment of the present invention makes further classification judgment when the registration duration is located in the registration duration interval corresponding to the two categories, thereby effectively making up for the deficiency that the current classification based only on the registration time may not accurately distinguish users, realizes the in-depth classification of users, improves the accuracy and adaptability of user classification, and also ensures the pertinence and effectiveness of subsequent user news push, thereby facilitating the improvement of user reading experience.

[0039] Further, the category tendency of the target user is determined, including: U1, extracting the access date of each record from the access information, and counting the access duration of the target user.

[0040] U2. Extract the browsing record log and interaction log under each record access date from the access information, and count the effective participation of the target user.

[0041] U3. Set the weights of access continuity and effective engagement, perform weighted sum calculation on the access continuity and effective engagement of target users, and use the calculation result as the active index value.

[0042] In a specific embodiment, news platforms usually hope that users can continue to visit and keep paying attention to the content to increase user retention. Therefore, the weight of access continuity can be set to a higher value, such as 60%, while effective participation can reflect the user's interest in news content and social interaction activity, and its weight can be set to a moderate value, such as 40%. The weights of access continuity and effective participation are 0.6 and 0.4 respectively.

[0043] U4. Match and compare the activity index value with the activity index value intervals corresponding to the set activity levels, and use the matched activity level as the target activity level, where the activity level is one of low, medium and high.

[0044] U5. If the categories in the judgment category group are category A and category B, the judgment category group will be labeled as Class I. If the categories are category B and category C, the judgment category group will be labeled as Class II.

[0045] U6. When the category group is marked as Category I, if the target activity level is low, Category A is used as the target user's tendency category; if the target activity level is medium or high, Category B is used as the target user's tendency category.

[0046] U7. When the category group is judged to be marked as Category II, if the target user's activity level is low, Category C is used as the target user's tendency category; if the target activity level is medium or high, Category B is used as the target user's tendency category.

[0047] The embodiment of the present invention analyzes the access continuity and participation effectiveness based on time as the baseline, and then confirms the activity level and tendency category analysis, which solves the shortcomings of traditional activity analysis that is only based on user login frequency and simple behavioral indicators. It fully reflects the actual participation of users on the platform, and can more accurately evaluate the user's activity level, thereby realizing personalized activity level classification and accurate user tendency category analysis.

[0048] Understandably, counting the access duration of the target user in step U1 includes: U1-1, marking each recorded access date on the electronic calendar, traversing each marked point, and obtaining the number of days between each marked point.

[0049] U1-2, with the number of the marked point as the horizontal axis and the number of days between visits as the vertical axis, construct a visit interval curve, and extract the slope from it as the interval change rate, which is recorded as k a .

[0050] It should be added that the slope of the extraction curve refers to the slope of the regression line corresponding to the extraction curve.

[0051] U1-3, if k0≤k a ≤k1, calculate the average of the interval days between each marked point, and use the calculated result as the target visit interval days, recorded as ΔD′, where k0 and k1 correspond to the lower limit and upper limit of the preset stable interval change rate range, respectively.

[0052] In a specific embodiment, when the interval change rate is 0, the surface interval days show a stable and regular change, that is, the guarantee of using the mean value to represent the target access interval days is higher, and the preset stable interval change rate interval is set based on the base point of 0. Specifically, k1 can be set to 0.1, and k0 can be set to -0.1.

[0053] U1-4, subtract ΔD′ from the set continuous defined visit interval number D0, and record the ratio of the difference to D0 as k ε , and used as the target user's access continuity.

[0054] U1-5, if k a >k1 or k a <k0, set the access continuous impact weight η.

[0055] Specifically, setting the access continuity impact weight includes: extracting the number of peak points and the difference in the number of days between each peak point and its adjacent valley point from the access interval curve, and recording the number of peak points as M.

[0056] Filter out the maximum value ΔD from the difference in days between each peak point and its adjacent valley point max , and at the same time, the average difference in the number of days between each peak point and its adjacent valley point is calculated, and the calculation result is recorded as

[0057] Will Target interval day difference ΔD σ , set the access continuous impact weight η, M′ is the number of reference fluctuation points, and e is a natural constant.

[0058] The embodiment of the present invention compensates for the error existing in the current analysis of using the mean as the target visit interval by setting the visit continuity impact weight, thereby more accurately and comprehensively evaluating the visit continuity impact, while increasing the reliability and reference value of the visit continuity analysis results.

[0059] It should be added that the specific value of M′ mainly depends on the number of annotation points. For example, when the number of annotation points is 30, M′ can be 3.

[0060] U1-6, when k a >k1, k ε *(1-η) is the target user’s access duration, when k a When k<k0, ε *(1+η) is the visit duration of the target user.

[0061] It needs to be explained that when k a When >k1, it indicates that the number of days between visits is increasing. When the average of the number of days between different marked points is used as the target number of days between visits, the calculated result of the visit duration will be greater than the actual result. Therefore, a correction is made to reduce k. ε *(1-η) is the target user's access duration, and when k a When <k0, it indicates that the number of days between visits is decreasing. When the average of the number of days between different marked points is used as the target number of days between visits, the calculated result of visit continuity will be less than the actual result. Therefore, an enlargement correction is performed, that is, k ε *(1+η) is the access duration of the target user.

[0062] It is also understandable that the effective participation of the target user is counted in step U2, including: U2-1, extracting the index value of each browsing index from the browsing record log, and constructing a browsing data table based on the index value of each browsing index under each record access date.

[0063] In a specific embodiment, the browsing index includes but is not limited to browsing news content and browsing time. For example, the index value of the browsing news content index can be the number of news contents browsed, and the index value of the browsing time index can be the cumulative browsing time.

[0064] U2-2. Compare the browsing data table with the browsing index items in the pre-set effective browsing evaluation data table.

[0065] It should be added that the pre-set effective browsing assessment data table is the same as the browsing indicators in the browsing data table, wherein the specific values ​​corresponding to each browsing indicator in the pre-set effective browsing assessment data table are pre-set effective browsing assessment thresholds or reference values, and the specific data source can be extracted from the user evaluation data table of the news platform.

[0066] U2-3. If the comparison relationship between a certain browsing indicator in the browsing data table and the effective browsing evaluation data table on a certain access date is greater than or equal to, the browsing indicator is marked as a valid indicator.

[0067] U2-4. Count the number of valid indicators and compare it with the number of browsing indicators. The ratio is recorded as the valid browsing indicator ratio. The valid browsing indicator ratio under each record access date is traversed.

[0068] U2-5. Record access dates with an effective browsing ratio greater than 0 and less than the set reference effective browsing index ratio as first-level effective access dates, and record access dates with an effective browsing ratio greater than the set reference effective browsing index ratio as second-level effective access dates.

[0069] U2-6. Count the number of first-level effective visit days and the number of second-level effective visit days, and compare them with the number of recorded visit dates respectively. The ratios are recorded as the first-level visit ratio and the second-level visit ratio, and the first-level visit ratio and the second-level visit ratio are standardized.

[0070] In a specific embodiment, the primary access ratio and the secondary access ratio are standardized to facilitate the operation of the statistical model, wherein the specific processing process of the standardization is: subtracting the primary access ratio from the set reference primary access ratio, comparing the difference with the reference primary access ratio, and taking the ratio as the excess of the primary access ratio.

[0071] Calculate the excess of the secondary access ratio in the same way as the excess of the first-level access ratio, set the weights of the first-level access ratio and the second-level access ratio, multiply the excess of the first-level access ratio and the second-level access ratio by their weights respectively, sum the calculation results, and use the sum as the standardized results of the first-level access ratio and the second-level access ratio.

[0072] U2-7. Use the standardized results as the independent variable of the statistical model, and use the dependent variable of the statistical model as the effective access degree.

[0073] In a specific embodiment, the statistical model may use a linear regression model and a logistic regression model for statistics. In order to ensure statistical efficiency, a linear regression model may be specifically selected for statistics.

[0074] U2-8. Extract the index value of each interaction index from the interaction log, and calculate the effective interaction degree in the same way as the analysis method of effective access degree.

[0075] In a specific embodiment, the interaction indicators include but are not limited to likes, favorites, comments and subscriptions. For example, the indicator value of the like indicator can be the number of likes, the indicator value of the favorite indicator can be the number of favorites, the indicator value of the comment indicator can be the number of comments, and the indicator value of the subscription indicator can be the number of subscriptions.

[0076] U2-9. Match and compare the effective access and effective interaction with the effective access interval and effective interaction interval corresponding to each set effective engagement, and use the matched effective engagement as the effective engagement of the target user.

[0077] It should be added that the matching effective engagement means that the effective access is within the effective access interval corresponding to the effective engagement, and the effective interaction is also within the effective interaction interval corresponding to the effective engagement.

[0078] Specifically, the benchmark recommendation model is modified in step S4, including: S4-1, extracting the marked source of each browsed news from the browsing record log, and the marked source includes retrieval, hot list and similar recommendation.

[0079] S4-2, integrating the marked sources of each browsed news under each recorded access date to obtain the marked sources of each cumulative browsed news.

[0080] S4-3. Compare the marked sources of each cumulatively viewed news, count the number of cumulatively viewed news whose marked sources are the hot list and similar recommendations, and compare them with the cumulative number of viewed news, and record the ratios as the hot adoption ratio k r and the recommended adoption ratio k t .

[0081] S4-4. Locate the benchmark weights for novel recommendations and similar recommendations from the benchmark recommendation model of the target user, and record them as weight 1 and weight 2 respectively.

[0082] S4-5. Determine the adjustment weight, confirm the adjustment value of the adjustment weight, and correct the corresponding weight in the benchmark recommendation model based on the adjustment value of the adjustment weight.

[0083] Further, the weight adjustment is determined and the adjustment value of the weight adjustment is confirmed, specifically including: X1, weight 1 is matched and compared with the novel recommendation weights corresponding to the preset adoption ratios, and the matched adoption ratio is used as the reference popular adoption ratio, recorded as k r ′, according to k r The setting method of ′ is similar to that of setting the reference recommendation ratio k t ′.

[0084] X2, will | k r -k r ′|>Δk is used as condition 1, Δk is the reference adoption ratio deviation, and |k t -k t ′|>Δk is used as condition 2.

[0085] In a specific embodiment, Δk may be specifically set to 0.1.

[0086] X3. If condition 1 is met, weight 1 is used as the adjustment weight, and the preset k r The corresponding novel recommendation weight is used as the adjustment value.

[0087] X4. If condition 2 is met, weight 2 is used as the adjustment weight, and the adjustment value under condition 2 is confirmed in the same way as the adjustment value under condition 1.

[0088] The embodiment of the present invention determines the adjustment weight by analyzing the browsing situation of news under different sources, and confirms the adjustment value of the adjustment weight, and then corrects the recommendation model, thereby avoiding the shortcomings of the current fixed model recommendation method. It can not only adapt to the changes in user interests in a timely manner, but also ensure the pertinence of the recommended content as much as possible, thereby effectively maintaining the balance between the personalization and novelty of the recommended content.

[0089] In yet another specific embodiment, the step of analyzing the target user's interest dynamic information in step S6 includes: extracting the marked topics of each browsed news from the browsing record log.

[0090] Based on the marked sources and marked topics of each browsed news on each record access date, a source list and a topic list are constructed, and the source change index values ​​and the topic change index values ​​are calculated based on them and used as interest dynamic information.

[0091] In a specific embodiment, the source list and the subject category list are constructed in the same manner, wherein a schematic source list under some record access dates is shown in Table 1.

[0092] Table 1 Illustrative source list of some records by access date

[0093]

[0094]

[0095] Among them, n represents the nth browsed news.

[0096] It should be added that the specific statistical process of calculating the source change index value is as follows: H1. Extract the number of browsed news of each marked source under each record access date from the source list.

[0097] H2. The ratio of the number of news viewed from each tagged source to the number of news viewed is taken as the browsing share of each tagged source.

[0098] H3. Take two record access dates as a group, construct each record access date group, and subtract the browsing proportions of each record access date in the same record access date group for the same tag source to obtain the browsing proportion difference of each record access date group corresponding to each tag source.

[0099] H4. Calculate the standard deviation of the browsing percentage difference of the same tag source in different record access date groups, obtain the calculated standard deviation value of each tag source, and use it as its degree of change.

[0100] H5. Compare the degree of change of each tag source with the set reference degree of change, and record the tag source with a degree of change greater than the set reference degree of change as the change source.

[0101] H6. Count the number of change sources and compare it with the number of marked sources to obtain the ratio of the number of change sources.

[0102] H7. Calculate the mean of the degree of change of each marked source, add the calculated result and the ratio of the number of change sources, and use the accumulated result as the source change index value.

[0103] It should also be added that the specific statistical process of calculating the value of the subject change index is as follows: extract each marked subject under each record access date from the subject list.

[0104] R1. Record the same marked topic as the same type of topic and count the number of types of marked topics.

[0105] R2. Calculate the standard deviation of the number of types of marked topics corresponding to each access date, and use the calculation result as the degree of change of topic types.

[0106] R3. Extract the number of browsed news of each tagged subject category under each record access date from the subject list, and calculate the degree of change of each tagged subject category in the same way as the degree of change calculation of each tagged source.

[0107] R4. Calculate the mean of the degree of change of each marked topic type, sum the calculated result with the degree of change of the topic type, and use the sum result as the topic change index value.

[0108] Furthermore, the setting of the monitoring adjustment frequency of the target user portrait in step S6 includes: recording the source change index value and the subject change index value as β0 and β1 respectively.

[0109] The comprehensive browsing change index value λ of the target user is calculated, λ=b1*β0+b2*β1, where b1 and b2 are the weights of the source change index and the theme change index respectively.

[0110] Match and compare λ with the comprehensive browsing change interval under each set monitoring adjustment frequency, and use the matched monitoring adjustment frequency as the monitoring adjustment frequency of the target user portrait.

[0111] It should be added that, from the perspective of change impact, the impact of source change on the recommendation model is greater than that of the topic, that is, the weight of the source change index is set greater than the weight of the topic change index. Specifically, b1 and b2 can be taken as 0.55 and 0.45 respectively.

[0112] The above contents are merely examples and explanations of the concept 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 concept of the invention or exceed the scope defined by the present invention, they shall all fall within the protection scope of the present invention.

Claims

1. A news content recommendation method based on multidimensional feature analysis, characterized in that: The method includes: Collect the target user's registration information and access information; Classify the target users according to their registration information and access information, and mark them by category; According to the category tag of the target user, a benchmark recommendation model matching the target user is selected from a preset benchmark recommendation model library; According to the access information of the target user, the baseline recommendation model is modified, and the modified baseline recommendation model is used as the target recommendation model for the target user; Extract personal information from the target user's registration information, associate the personal information, category label and target recommendation model to form a target user profile; Analyze the interest dynamics of target users and set the monitoring and adjustment frequency of target user portraits; Upload the target user portrait, the monitoring adjustment frequency of the target user portrait, and the target recommendation model to the news push platform; The classifying the target user includes: Based on the pre-set registration duration intervals corresponding to each category, the categories are sorted from small to large according to their registration durations, and are recorded as category A, category B, and category C in sequence; Extract the registration duration from the target user’s registration information; If the target user's registration duration is within the registration duration range of a certain category, then the category is used as the target user's category; If the registration duration of the target user is at the intersection of the registration duration interval of a certain category and the registration duration interval of another category, the two categories are combined into a judgment category group, the category tendency of the target user is judged, and then the tendency category is used as the category of the target user; The determining of the category tendency of the target user includes: Extract the access date of each record from the access information and count the access duration of the target user; Extracting browsing record logs and interaction logs on each recorded access date from the access information, and counting the effective participation of the target user; Set the weights of access continuity and effective engagement, calculate the weighted sum of the target user's access continuity and effective engagement, and use the calculation result as the active index value; Matching and comparing the activity index value with the activity index value intervals corresponding to the set activity levels, and using the matched activity level as the target activity level, where the activity level is one of low, medium and high; If the categories in the judgment category group are category A and category B, the judgment category group is labeled as category I; if the categories are category B and category C, the judgment category group is labeled as category II; When the category group is marked as Category I, if the target activity level is low, Category A is used as the target user's tendency category; if the target activity level is medium or high, Category B is used as the target user's tendency category; When the category group is marked as Category II, if the target user's activity level is low, Category C is used as the target user's tendency category; if the target activity level is medium or high, Category B is used as the target user's tendency category; The modifying of the benchmark recommendation model includes: Extracting the marked source of each browsed news from the browsing record log, wherein the marked source includes search, hot list and similar recommendation; Integrate the marked sources of each browsed news under each record access date to obtain the marked sources of each cumulative browsed news; Compare the marked sources of each cumulatively viewed news, count the cumulative number of viewed news whose marked sources are hot lists and similar recommendations, and compare them with the cumulative number of viewed news, and record the ratio as the hot adoption ratio Compared with the recommended adoption ; Locate the benchmark weights for novel recommendations and similar recommendations from the benchmark recommendation model of the target user, denoted as weight 1 and weight 2 respectively; Determine the adjustment weights, confirm the adjustment values ​​of the adjustment weights, and correct the corresponding weights in the benchmark recommendation model based on the adjustment values ​​of the adjustment weights; The determining of the adjustment weight and confirming the adjustment value of the adjustment weight specifically includes: The weight 1 is matched and compared with the novel recommendation weights corresponding to the preset adoption ratios, and the matched adoption ratio is used as the reference popular adoption ratio, which is recorded as ,according to The setting method is similar to the setting of reference recommendation adoption ratio ; Will As condition 1, is the reference adoption ratio deviation, As condition 2; If condition 1 is met, weight 1 is used as the adjustment weight, and the preset The corresponding novel recommendation weight is used as the adjustment value; If condition 2 is met, weight 2 is used as the adjustment weight, and the adjustment value under condition 2 is confirmed in the same way as the adjustment value under condition 1.

2. The news content recommendation method based on multidimensional feature analysis as claimed in claim 1, characterized in that: The statistical target user's access duration includes: Mark each record access date on the electronic calendar, traverse each marked point, and obtain the number of days between each marked point; The visit interval curve is constructed with the number of the marked point as the horizontal axis and the number of days between visits as the vertical axis, and the slope is extracted from it as the interval change rate, which is recorded as ; like , calculate the average of the interval days between each marked point, and use the calculated result as the target visit interval days, recorded as ,in, and They correspond to the lower limit and upper limit of the preset stable interval change rate range respectively; Will Set the number of days between visits to define the duration Subtract the difference and The ratio of , and as the target user's access continuity; like or , set the visit continuous impact weight ; when When As the target user's access duration, When The duration of target users’ visits.

3. A news content recommendation method based on multidimensional feature analysis as claimed in claim 2, characterized in that: The settings access continues to affect weights, including: From the access interval curve, the number of peak points and the difference in days between each peak point and its adjacent valley point are extracted, and the number of peak points is recorded as ; Filter out the maximum value from the difference in days between each peak point and its adjacent valley point , and at the same time, the average difference in the number of days between each peak point and its adjacent valley point is calculated, and the calculation result is recorded as ; Will Target interval days difference , set the visit continuous impact weight , , To set the number of reference fluctuation points, is a natural constant.

4. The news content recommendation method based on multidimensional feature analysis as claimed in claim 1, characterized in that: The statistical target user's effective participation degree includes: Extracting the index value of each browsing index from the browsing record log, and constructing a browsing data table based on the index value of each browsing index under each record access date; Compare the browsing data table with the browsing index items in the preset effective browsing evaluation data table; If the comparison relationship between a browsing indicator in the browsing data table and the effective browsing evaluation data table on a certain access date is greater than or equal to, the browsing indicator is marked as a valid indicator; Count the number of valid indicators and compare them with the number of browsing indicators. The ratio is recorded as the valid browsing indicator ratio. The valid browsing indicator ratio under each record access date is traversed. Record the access date of the record whose effective browsing ratio is greater than 0 and less than the set reference effective browsing index ratio as the first-level effective access date, and record the access date of the record whose effective browsing ratio is greater than the set reference effective browsing index ratio as the second-level effective access date; Count the number of first-level effective visit days and the number of second-level effective visit days, and compare them with the number of recorded visit dates respectively. The ratios are recorded as the first-level visit ratio and the second-level visit ratio, and the first-level visit ratio and the second-level visit ratio are standardized; The standardized results are used as the independent variables of the statistical model, and the dependent variable of the statistical model is used as the effective access degree; Extract the index value of each interaction index from the interaction log, and calculate the effective interaction degree in the same way as the analysis method of effective access degree; The effective access and effective interaction are matched and compared with the effective access interval and effective interaction interval corresponding to each set effective engagement, and the matched effective engagement is used as the effective engagement of the target user.

5. The news content recommendation method based on multidimensional feature analysis as claimed in claim 1, characterized in that: The analyzing the target user's interest dynamic information includes: Extracting the marked topics of each browsed news from the browsing record log; Based on the marked sources and marked topics of each browsed news on each record access date, a source list and a topic list are constructed, and the source change index values ​​and the topic change index values ​​are calculated based on them and used as interest dynamic information.

6. A news content recommendation method based on multidimensional feature analysis as claimed in claim 5, characterized in that: The step of setting the monitoring adjustment frequency of the target user portrait includes: The source change index value and the topic change index value are respectively recorded as and ; Count the comprehensive browsing change index values ​​of target users , , and are the weights of the source change index and the topic change index, respectively; Will The monitoring adjustment frequency is matched and compared with the comprehensive browsing change interval under each set monitoring adjustment frequency, and the matched monitoring adjustment frequency is used as the monitoring adjustment frequency of the target user portrait.