Personalized content intelligent recommendation method and system based on digital financial AI platform

Abstract

The present application relates to the field of content recommendation, more particularly, the present application relates to a personalized content intelligent recommendation method and system based on a digital financial AI platform, the method comprising: obtaining articles related to financial objects held by a user, extracting relevant words, and calculating the price synchronization coefficient of each relevant word in a continuous time unit; constructing a price change interval for the price rise or fall of the object, integrating the price synchronization coefficient of each word in the interval, and removing the relevant word with the smallest integral value through iteration, stopping when the comprehensive influence degree of the remaining word set reaches a maximum value, obtaining an effective relevant word set and updating the word evaluation thereof; calculating a recommendation index according to the word evaluation and proportion of the effective relevant words contained in the article to be recommended, and pushing the article to the user if the recommendation index exceeds a threshold value. The present application can effectively improve the accuracy, timeliness and personalization degree of financial information recommendation, and help users accurately capture key market information.

Description

Technical Field

[0001] This invention relates to the field of content recommendation. More specifically, this invention relates to a personalized content intelligent recommendation method and system based on a digital financial AI platform. Background Technology

[0002] Digital financial platforms generate massive amounts of news, research reports, announcements, and social media text daily. The prices of users' holdings (such as stocks, futures, and funds) are driven by multiple factors, including macroeconomic conditions, industry dynamics, and market sentiment. Changes in this information can affect the prices of these assets immediately or proactively. To help users quickly extract highly relevant content from this vast amount of information, existing technologies typically employ recommendation methods based on user profiles or collaborative filtering.

[0003] However, these traditional methods have significant limitations: static profiles struggle to reflect the dynamic relationship between a user's holdings and real-time information; collaborative filtering methods rely on group behavior patterns, making it difficult to capture the real-time, non-linear relationship between individual holdings and specific information. Therefore, recommended content often lags behind or lacks strong relevance, failing to meet users' urgent need for accurate and timely information. Summary of the Invention

[0004] To address the technical problem that traditional recommendation methods are limited in meeting users' urgent need for accurate and timely information, this invention provides solutions in the following aspects.

[0005] In the first aspect, the personalized content intelligent recommendation method based on the digital finance AI platform includes:

[0006] The system acquires articles published within a preset time period that are related to the financial assets held by the user, constructs a keyword article library, and extracts multiple related keywords; it calculates the price synchronization coefficient of each related keyword in a continuous time unit, which is calculated from the changes in the price of the underlying asset and the frequency of occurrence of the related keyword.

[0007] Construct a price fluctuation range for the target price to rise or fall within a preset time period. For each price fluctuation range, obtain all relevant words appearing in the range, integrate the price synchronization coefficient of each word within the range to obtain the integral value, iteratively eliminate the relevant words corresponding to the minimum integral value, and calculate the comprehensive influence of the remaining relevant words on the price fluctuation after each elimination. When the comprehensive influence reaches the maximum value, stop eliminating, determine the current remaining relevant words as the set of effective relevant words, and update the word evaluation of each effective relevant word.

[0008] For articles to be recommended, a recommendation index is calculated based on the word evaluation and proportion of effective related words contained therein; when the recommendation index exceeds a preset threshold, the article is pushed to users who hold the corresponding target.

[0009] Preferably, the construction of the keyword article database includes:

[0010] The importance of the target as a keyword in each article is calculated, and the importance is determined based on the word frequency ratio of the keyword in the article and its evenness of distribution in each paragraph of the article;

[0011] Articles with an importance level below a preset threshold are excluded to form the keyword article library.

[0012] Preferably, the extraction of related words includes:

[0013] Based on part-of-speech tagging, at least one of nouns, verbs, and adjectives is extracted from the article, and words without substantive meaning in the stop word list are filtered out.

[0014] Preferably, obtaining the price synchronization coefficient specifically includes:

[0015] For a single related term, under continuous time units, the absolute value of the difference between the price of the target in the current time unit and the price in the previous time unit is calculated as the ratio of the price in the previous time unit to the price in the current time unit, which is used as the first ratio.

[0016] In the keyword article database, identify all articles published within the current time unit and mark them as target articles. Count the total number of occurrences of the relevant keyword in these target articles and calculate the average number of occurrences of the relevant keyword in each target article.

[0017] The ratio of the average occurrence count to the total number of target articles is calculated as the second ratio;

[0018] The product of the first ratio and the second ratio is used as the price synchronization coefficient of the related term in the current time unit.

[0019] Preferably, the price fluctuation range of the target asset within a preset time period, where the price rises or falls overall, includes:

[0020] Curve fitting is performed on the price series of the target asset formed under all time units within the preset time period. The continuously monotonically rising curve segment is extracted as the overall price fluctuation range, and the continuously monotonically falling curve segment is extracted as the overall price fluctuation range.

[0021] Preferably, the acquisition of the overall impact level includes:

[0022] The relative change in the price of the target within the interval at the start and end times is calculated as the first factor; the ratio of the number of currently remaining related words after elimination to the number of all related words before the iteration is calculated as the second factor; the sum of the integral values ​​of the currently remaining related words within the interval after elimination is calculated and the sum of the integral values ​​is transformed by the hyperbolic tangent function, and the result of the transformation is used as the third factor.

[0023] The product of the calculated first, second, and third factors is taken as the overall degree of influence.

[0024] Preferably, the updating of the word evaluation of each valid related word includes:

[0025] Calculate the average of the overall influence of each related word when it is determined to be a valid related word across all price fluctuation ranges, and use this as the word evaluation for that related word.

[0026] Preferably, the calculation process of the recommendation index includes:

[0027] For any article in the aforementioned keyword article database, the ratio of the number of occurrences of each effective related word in that article to the total number of occurrences of all effective related words in that article is calculated as the third ratio;

[0028] The recommendation index is obtained by summing the products of the third ratio and the word evaluations of all valid related words.

[0029] Secondly, a personalized content intelligent recommendation system based on a digital financial AI platform includes: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the personalized content intelligent recommendation method based on the digital financial AI platform described in any one of the above claims is implemented.

[0030] The beneficial effects of this invention are:

[0031] This invention achieves a dynamic correlation between financial information and market fluctuations by constructing a price synchronization coefficient that integrates real-time market data and text data. Furthermore, through price range division, point accumulation, and iterative elimination mechanisms, it accurately selects a set of effective relevant words that drive prices, significantly improving the information signal-to-noise ratio. Finally, based on dynamic word evaluation and personalized recommendation index, it pushes highly relevant and timely customized content to users, effectively overcoming the shortcomings of traditional recommendation methods such as lag, generalization, and information overload, and providing a precise, adaptive, and explainable intelligent recommendation solution for digital financial platforms. Attached Figure Description

[0032] Figure 1 This is a flowchart of steps S1-S4 in the personalized content intelligent recommendation method based on a digital financial AI platform according to an embodiment of the present invention.

[0033] Figure 2 This is a schematic diagram of the structure of a personalized content intelligent recommendation system based on a digital financial AI platform, according to an embodiment of the present invention. Detailed Implementation

[0034] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0035] This invention is mainly applied in the field of digital finance, aiming to provide highly personalized intelligent content recommendation services for users who hold specific financial assets (such as stocks, funds, etc.).

[0036] Reference Figure 1 The personalized content intelligent recommendation method based on the digital finance AI platform includes steps S1-S4, as follows:

[0037] S1: Obtain articles published within a preset time period that are related to the financial assets held by the user, build a keyword article library, and extract multiple related keywords.

[0038] Digital financial platforms generate massive amounts of text information daily, but not all articles mentioning a user's held assets are highly relevant or have a substantial impact. Traditional recommendation methods rely on fixed keywords or static user profiles, failing to dynamically assess the true strength of the correlation between individual articles and the underlying assets, and are prone to introducing noise and lowering the signal-to-noise ratio. To accurately select articles that truly discuss the underlying assets, it is necessary to build a high-quality keyword article library and extract keywords (related words) that may affect prices, laying the foundation for subsequent quantitative analysis.

[0039] The construction process of the above-mentioned subject article database is generally as follows:

[0040] First, use any financial asset held by the user (such as a stock code, fund name, etc.) as the keyword.

[0041] Then, for each article containing the keyword published within a preset time period, such as the last 24 hours or one week, the original articles are further preprocessed, including: removing HTML / XML tags, irrelevant symbols, and garbled characters; splitting the text into ordered sentence sequences; and then dividing the article into several paragraphs according to paragraph markers (such as line breaks) or natural sentence segmentation rules. Subsequently, using Chinese word segmentation tools or English word segmentation and lemmatization techniques, each paragraph is segmented to obtain the basic vocabulary unit sequence that constitutes the article, and the total number of words in the article (i.e., the sum of vocabulary units in all paragraphs) is counted. At the same time, the sequence and number of vocabulary units in each paragraph are recorded.

[0042] Next, the importance of the keyword in each article is calculated.

[0043] Taking a single article as an example, the formula for calculating the importance level mentioned above is as follows:

[0044]

[0045] In the formula, To determine the importance of the aforementioned keywords in a single article, This represents the percentage of the total number of words related to this keyword in the article. It is an exponential function with the natural number e as its base. This is the set of the number of times the keyword appears in each paragraph of the article. The standard deviation is denoted as .

[0046] The above calculations take into account both the frequency percentage of keywords and the uniformity of their distribution in the article. A high value indicates that the article revolves around the keywords. A high value indicates that the keywords are evenly distributed across paragraphs, and that the article is coherent and focused on a single theme.

[0047] Finally, an importance threshold is set, and articles with an importance lower than this threshold are excluded. The remaining articles constitute the keyword article library. This keyword article library is a dynamic collection that is continuously updated over time.

[0048] After obtaining the subject term article database, relevant words are extracted from all articles in the database. Natural language processing techniques are used for part-of-speech tagging to extract words with meaning, such as nouns, verbs, and adjectives, while words without meaning, such as prepositions, conjunctions, and auxiliary words, from the stop word list are filtered out, forming an initial set of relevant words.

[0049] S2: Calculate the price synchronization coefficient of each related term over a continuous time unit.

[0050] It is important to consider that the impact of market information on prices is often time-sensitive. If the frequency of a word's occurrence is highly synchronized with the price fluctuations of the underlying asset, it is likely to reflect the immediate information or market sentiment driving price changes. Traditional methods lack the ability to quantify the "word frequency-price" linkage at such micro-timescales (e.g., seconds). To identify relevant words that are truly synchronized with price fluctuations, it is necessary to calculate the price synchronization coefficient of each relevant word over continuous time units.

[0051] In one embodiment, the preset time period in S1 is divided into continuous basic time units such as seconds (which can be adjusted according to computing power). Additionally, through the market data interface integrated into the platform (such as the API of a stock exchange or financial data service provider), the price sequence of the target asset sampled at fixed time units within the same preset time period is synchronously obtained. This sequence is an ordered array, with each element containing a timestamp and the corresponding price (such as closing price or latest transaction price).

[0052] For each related word in the related word set finally obtained in S1 above, calculate its price synchronization coefficient in each time unit.

[0053] Taking a single related term as an example, the calculation process for the price synchronization coefficient mentioned above is generally as follows:

[0054] Assuming the current time unit is The previous time unit of the current time unit is , respectively obtain the first and the The price of the object at the end of each time unit, and then the price of the object in the [number]th time unit. The price per unit of time and the price at the time unit The absolute value of the difference between prices in the time units and the in the time unit The ratio of prices in adjacent time units is used as the first ratio. This first ratio represents the rate of price change between adjacent time units; the larger the value, the more volatile the price fluctuations.

[0055] In the Within a given time unit, the average number of times the relevant keyword appeared in all articles in the keyword article database was calculated, and the number of times the keyword appeared in the first article was also calculated. The total number of articles containing the relevant keyword in the keyword article database within a given time unit.

[0056] Calculate in the first Within a given time unit, the ratio of the average number of times a related word appears in all articles in the keyword article database to the total number of articles in the keyword article database containing that related word is used as the second ratio. This second ratio represents the frequency density of the related word in articles during the same period; a higher value indicates that the related word is more prominent in the current information flow.

[0057] Furthermore, the product of the first ratio and the second ratio calculated above is taken as the value of the related word in the first... The price synchronization coefficient at each time unit comprehensively reflects the intensity of the synchronous occurrence of the relevant word in the relevant text at the moment when prices fluctuate.

[0058] Based on the above operations, the price synchronization coefficient of all related terms at any given time unit can be obtained similarly.

[0059] S3: Construct a price fluctuation range for the target asset's price to rise or fall within a preset time period. For each price fluctuation range, obtain all relevant words appearing within the range, integrate the price synchronization coefficient of each word within the range to obtain the integral value, iteratively eliminate the relevant words corresponding to the minimum integral value, and calculate the comprehensive influence of the remaining relevant words on the price fluctuation after each elimination. When the comprehensive influence reaches the maximum value, stop eliminating, determine the current remaining relevant words as the set of effective relevant words, and update the word evaluation of each effective relevant word.

[0060] It is also important to consider that price fluctuations are often driven by information or sentiment jointly represented by multiple related words over a period of time, rather than by a single word. Traditional methods cannot assess the synergistic effect of multiple words within the complete range of a price trend (upward or downward wave) and tend to retain high-frequency but irrelevant noise words. Therefore, it is necessary to first identify meaningful price movement ranges (upward or downward waves), then quantify the overall contribution of each word within that range through integration, and finally use an iterative elimination mechanism to select the most explanatory and effective combination of related words for this price movement.

[0061] In one embodiment, a price fluctuation range is first constructed: for the price series of the underlying asset within the aforementioned preset time period, a curve fitting algorithm, such as cubic spline interpolation or polynomial fitting, is used to smooth the discrete price data, generating a continuous fitted curve. This fitted curve is then analyzed, and all continuously monotonically rising curve segments are extracted as the overall upward price fluctuation range, while all continuously monotonically falling curve segments are extracted as the overall downward price fluctuation range.

[0062] Next, for a specific price fluctuation range, all relevant words that have appeared in all time units within that price fluctuation range are obtained. For each relevant word, its price synchronization coefficient is integrated (definite integral) within that price fluctuation range to obtain the integral value of each relevant word. This integral value represents the overall correlation strength between the relevant word and price fluctuations throughout the entire price range.

[0063] After obtaining the integral values ​​of all relevant terms within the price fluctuation range, iterative elimination is performed based on these integral values. The general operation of iterative elimination is as follows:

[0064] First, all relevant words contained within the price fluctuation range are used as the initial set of relevant words.

[0065] Then, for the initial set of related words, the first elimination begins, that is, the related words with the smallest integral value in the initial set of related words are eliminated, and the remaining related words are used as the new set of related words.

[0066] For a new set of related terms, calculate the overall impact of its influence on price changes:

[0067] Calculate the relative change in the price of the underlying asset within the price fluctuation range at the start and end times (i.e., the ratio of the difference between the price of the underlying asset at the end of the price fluctuation range and the price of the underlying asset at the start to the price of the underlying asset at the start). Use this relative change as the first factor to represent the overall scale of price fluctuations within this price fluctuation range.

[0068] The ratio of the number of related words in the new set of related words to the number of related words in the initial set of related words is calculated as a second factor, representing the retention rate of the remaining related words.

[0069] Calculate the sum of the integral values ​​of all related words in the new set of related words, and perform a hyperbolic tangent function transformation on the sum of integral values. Use the result of the transformation as the third factor.

[0070] Finally, the product of the first, second, and third factors is used as the overall degree of influence.

[0071] After calculating the overall impact after the first elimination, the second elimination is started, and the above elimination operation and the operation of calculating the overall impact are performed. The third, fourth and so on are performed in sequence until the overall impact reaches the first maximum value. Then the iterative elimination stops, and the remaining related words are formed into a set of effective related words.

[0072] Furthermore, iterate through all price fluctuation ranges. For any related word, find all price fluctuation ranges that are considered valid related words, and calculate the average of the overall influence of these price fluctuation ranges as the word evaluation of the related word.

[0073] The S3 operation described above ensures the accuracy and timeliness of price fluctuation range construction through explicit price data access and preprocessing. Subsequent focus on complete trending price waves, employing integration and iterative elimination strategies, filters out a large number of randomly occurring or weakly contributing noise words, accurately identifying the word combinations that truly function during key price movements. Dynamically updated word evaluations assign weights based on historical effectiveness to each relevant word, enabling the model to learn cumulatively and distinguish between long-term effective factors and short-term incidental words.

[0074] S4: For articles to be recommended, calculate the recommendation index based on the word evaluation and proportion of the effective related words contained therein; when the recommendation index exceeds the preset threshold, push the article to users who hold the corresponding target.

[0075] The ultimate goal of this invention is to recommend the most valuable articles to users. The recommendation value of an article depends on whether its content contains words that currently have a significant impact on the price of the user's holdings, and the prominence of these words within the text. By combining the article content with dynamically calculated effective relevant words and their historical effectiveness (word evaluation), a quantitative recommendation index can be calculated, enabling objective and personalized push decisions.

[0076] In one embodiment, for any article in the keyword article library or a newly crawled article to be recommended, all effective related words contained in the article are identified. For each effective related word, the number of times the related word appears in the article is counted, as well as the total number of times all effective related words appear in the article. Then, the ratio of the number of times a single effective related word appears in the article to the total number of times all effective related words appear is calculated as a third ratio. The sum of the products of the third ratio and the word evaluation of all effective related words is calculated to obtain the recommendation index of the article.

[0077] Similarly, the recommendation index of all articles in the keyword article library or newly crawled articles to be recommended can be obtained by following the above operation.

[0078] Furthermore, a threshold is preset, which can be determined in one or more of the following ways:

[0079] The first method involves collecting recommendation indices for all articles calculated for the target within a historical time period, forming a historical distribution. Based on this historical data distribution, a threshold is determined as a certain statistical percentile, which can be selected from the 70th, 80th, or 90th percentile to balance detection sensitivity and false alarm rate.

[0080] The second approach is to set an absolute threshold based on experience, such as 0.5.

[0081] The third approach is to adjust the threshold based on the user's risk preference or information density needs. A lower threshold can be set for users who want more comprehensive information, while a higher threshold can be set for users who prefer concise information.

[0082] The recommendation index calculated above is compared with the final threshold. If the recommendation index is greater than or equal to the threshold, the corresponding article is determined to have high recommendation value for the user, and the push process is triggered; otherwise, it is not pushed.

[0083] This invention also provides a personalized content intelligent recommendation system based on a digital finance AI platform. For example... Figure 2 As shown, the system includes a processor and a memory. The memory stores computer program instructions. When the computer program instructions are executed by the processor, the personalized content intelligent recommendation method based on a digital financial AI platform according to the first aspect of the present invention is implemented.

[0084] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0085] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A personalized content intelligent recommendation method based on a digital financial AI platform, characterized in that, include: Retrieve articles published within a preset time period that are related to the financial assets held by the user, build a keyword article library, and extract multiple related keywords; Calculate the price synchronization coefficient for each related term over a continuous time unit. This price synchronization coefficient is calculated from the price change of the underlying asset and the frequency of occurrence of the related term. Construct a price fluctuation range for the target price to rise or fall within a preset time period. For each price fluctuation range, obtain all relevant words appearing in the range, integrate the price synchronization coefficient of each word within the range to obtain the integral value, iteratively eliminate the relevant words corresponding to the minimum integral value, and calculate the comprehensive influence of the remaining relevant words on the price fluctuation after each elimination. When the comprehensive influence reaches the maximum value, stop eliminating, determine the current remaining relevant words as the set of effective relevant words, and update the word evaluation of each effective relevant word. The acquisition of the overall impact level includes: The relative change in the price of the target within the interval at the start and end times is calculated as the first factor; the ratio of the number of currently remaining related words after elimination to the number of all related words before the iteration is calculated as the second factor; the sum of the integral values ​​of the currently remaining related words within the interval after elimination is calculated and the sum of the integral values ​​is transformed by the hyperbolic tangent function, and the result of the transformation is used as the third factor. The product of the calculated first factor, second factor, and third factor is taken as the overall degree of influence. The word evaluation of each effective related word is updated as follows: Calculate the average of the overall influence of each related word when it is determined to be a valid related word across all price fluctuation ranges, and use this as the word evaluation of the related word; For articles to be recommended, a recommendation index is calculated based on the word evaluation and proportion of effective related words contained therein; when the recommendation index exceeds a preset threshold, the article is pushed to users who hold the corresponding target. 2.The personalized content intelligent recommendation method based on the digital financial AI platform according to claim 1, characterized in that, The construction of the subject article database includes: The importance of the target as a keyword in each article is calculated, and the importance is determined based on the word frequency ratio of the keyword in the article and its evenness of distribution in each paragraph of the article; Articles with an importance level below a preset threshold are excluded to form the keyword article library. 3.The personalized content intelligent recommendation method based on the digital financial AI platform according to claim 1, characterized in that, The extraction of related words includes: Based on part-of-speech tagging, at least one of nouns, verbs, and adjectives is extracted from the article, and words without substantive meaning in the stop word list are filtered out. 4.The personalized content intelligent recommendation method based on the digital financial AI platform according to claim 1, characterized in that, The acquisition of the price synchronization coefficient specifically includes: For a single related term, under continuous time units, the absolute value of the difference between the price of the target in the current time unit and the price in the previous time unit is calculated as the ratio of the price in the previous time unit to the price in the current time unit, which is used as the first ratio. In the keyword article database, identify all articles published within the current time unit and mark them as target articles. Count the total number of times the related word appears in these target articles and calculate the average number of times the related word appears in each target article. The ratio of the average occurrence count to the total number of target articles is calculated as the second ratio; The product of the first ratio and the second ratio is used as the price synchronization coefficient of the related term in the current time unit. 5.The personalized content intelligent recommendation method based on the digital financial AI platform according to claim 1, characterized in that, The price fluctuation range of the target asset within a preset time period, including overall increase or decrease, includes: Curve fitting is performed on the price series of the target asset formed under all time units within the preset time period. The continuously monotonically rising curve segment is extracted as the overall price fluctuation range, and the continuously monotonically falling curve segment is extracted as the overall price fluctuation range. 6.The personalized content intelligent recommendation method based on the digital financial AI platform according to claim 1, characterized in that, The calculation process for the recommendation index includes: For any article in the aforementioned keyword article database, the ratio of the number of occurrences of each effective related word in that article to the total number of occurrences of all effective related words in that article is calculated as the third ratio; The recommendation index is obtained by summing the products of the third ratio and the word evaluations of all valid related words.

7. A personalized content intelligent recommendation system based on a digital financial AI platform, characterized in that: include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the personalized content intelligent recommendation method based on a digital financial AI platform according to any one of claims 1-6.

Images