Personalized content recommendation methods and systems for financial information

Abstract

This application provides a method and system for personalized content recommendation of financial information. The method includes: performing semantic parsing, event normalization, and cross-source association processing on financial information data to generate event impact units corresponding to financial events; extracting time-series fluctuation features and topic evolution features based on market state data to generate market state representations corresponding to each event impact unit; extracting long-term interest features and conversation intent features based on user interaction data to construct decision state representations; establishing causal association estimation results including exposure bias and timeliness bias, and determining a set of candidate event impact units; and performing distribution calibration sorting and intra-cluster mapping processing under confidence constraints based on the candidate event impact unit set, market state representations, and decision state representations to generate personalized financial information recommendation results for the target user. This application can reduce duplicate recommendation rates, decrease recommendation bias, and improve timeliness adaptability and personalization accuracy.

Description

Technical Field

[0001] This application relates to the field of Internet technology, and in particular to a personalized content recommendation method and system for financial information. Background Technology

[0002] With the development of internet information platforms, intelligent recommendation technologies, and financial information service technologies, personalized content distribution for financial information has become an important technological direction for improving information reach efficiency and user service capabilities. This type of technology typically targets financial information such as stocks, funds, bonds, macroeconomics, and industry trends. It models and analyzes massive amounts of information content and combines this with user interaction behaviors such as browsing history, clicks, dwell time, and subscriptions to recommend relevant financial information to different users.

[0003] In existing technologies, common solutions often employ recommendation methods based on keyword matching, collaborative filtering, content semantic analysis, or user behavior sequence modeling. These methods perform correlation calculations between financial news text and user interests to generate recommendation results. Some solutions also incorporate news popularity, publication time, or tag information to optimize the ranking of recommendation results.

[0004] However, financial news is characterized by strong event-driven nature, rapid timeliness changes, high repetition in cross-source reporting, and close correlation with market conditions. Existing technologies typically recommend individual news articles, lacking integrated processing of hierarchical semantics of financial matters, which easily leads to repeated recommendations of similar events. At the same time, existing technologies do not adequately consider bias factors such as exposure position, news popularity, and timeliness decay, causing recommendation results to deviate from users' actual focus. Furthermore, existing technologies have failed to effectively combine changes in market conditions with users' short-term decision-making intentions to collaboratively model the recommendation process, resulting in shortcomings in the timeliness matching, reasonable topic distribution, and personalized adaptation of recommendation results. Summary of the Invention

[0005] In view of this, embodiments of this application provide a personalized content recommendation method and system for financial information to solve the problems of repeated recommendations of similar events, large recommendation bias, and insufficient timeliness matching and personalization adaptation in the prior art.

[0006] A first aspect of this application provides a personalized content recommendation method for financial information, comprising: acquiring financial information data, market status data, and user interaction data corresponding to a target user; performing semantic parsing, event normalization, and cross-source association processing on the financial information data to generate event impact units corresponding to financial matters; extracting time-series fluctuation features and topic evolution features based on the market status data to generate market status representations corresponding to each event impact unit; extracting long-term interest features and conversation intent features based on the user interaction data, and constructing a decision-making state representation corresponding to the target user in conjunction with the event impact units; establishing a causal association estimation result including exposure bias and timeliness bias based on the event impact units, market status representation, and decision-making state representation, and determining a set of candidate event impact units based on the causal association estimation result; and performing distribution calibration sorting and intra-cluster mapping processing under confidence constraints based on the set of candidate event impact units, market status representation, and decision-making state representation to generate personalized financial information recommendation results corresponding to the target user.

[0007] A second aspect of this application provides a personalized content recommendation system for financial information, comprising: an acquisition module, configured to acquire financial information data, market status data, and user interaction data corresponding to a target user; perform semantic parsing, event normalization, and cross-source association processing on the financial information data to generate event impact units corresponding to financial matters; an extraction module, configured to extract time-series fluctuation features and topic evolution features based on market status data to generate market status representations corresponding to each event impact unit; a construction module, configured to extract long-term interest features and conversation intent features based on user interaction data, and construct a decision state representation corresponding to the target user in conjunction with the event impact units; a determination module, configured to establish a causal association estimation result including exposure bias and timeliness bias based on the event impact units, market status representations, and decision state representations, and determine a set of candidate event impact units based on the causal association estimation result; and a generation module, configured to perform distribution calibration sorting and intra-cluster mapping processing under confidence constraints based on the set of candidate event impact units, market status representations, and decision state representations to generate personalized financial information recommendation results corresponding to the target user.

[0008] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0009] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0010] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: By acquiring financial information data, market status data, and user interaction data corresponding to target users, this application performs semantic parsing, event normalization, and cross-source association processing on the financial information data to generate event impact units corresponding to financial matters. Based on market status data, it extracts time-series fluctuation features and theme evolution features to generate market status representations corresponding to each event impact unit. Based on user interaction data, it extracts long-term interest features and conversation intent features, and combines these with event impact units to construct decision-making status representations for target users. Based on event impact units, market status representations, and decision-making status representations, it establishes causal association estimation results including exposure bias and timeliness bias, and determines a set of candidate event impact units based on the causal association estimation results. Based on the candidate event impact unit set, market status representations, and decision-making status representations, it performs distribution calibration sorting and intra-cluster mapping processing under confidence constraints to generate personalized financial information recommendation results for target users. This application can reduce duplicate recommendation rates, decrease recommendation bias, and improve timeliness adaptability and personalization accuracy. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating the personalized content recommendation method for financial information provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of a personalized content recommendation system for financial information provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0013] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0014] In existing technologies, personalized recommendations for financial information typically involve modeling the information text, tag information, and user behavior data such as historical clicks, browsing, and dwell time. Recommendation results are then generated based on keyword matching, content semantic analysis, collaborative filtering, or behavioral sequence analysis. Some solutions further rank the recommendation results by incorporating information publication time, popularity, or category attributes to improve information distribution efficiency and user content reach.

[0015] However, financial news is characterized by strong event-driven nature, rapid timeliness changes, high repetition across sources, and close correlation with market conditions. Existing technologies mostly still use individual news articles as recommendation targets, lacking consolidated modeling and unified representation of the same financial event, easily leading to duplicate recommendations of similar events. Furthermore, existing technologies do not adequately consider recommendation biases caused by factors such as exposure position, news popularity, and time decay, causing recommendations to deviate from users' actual concerns. In addition, existing technologies fail to co-model changes in market conditions with users' short-term decision-making intentions, thus remaining insufficient in terms of timeliness matching and personalized adaptability.

[0016] In view of the aforementioned problems in existing technologies, this application proposes a personalized content recommendation method for financial information. The method first acquires financial information data, market status data, and user interaction data. Semantic parsing, event normalization, and cross-source association processing are performed on the financial information data to generate event impact units corresponding to financial matters, thereby elevating the recommendation object from a single piece of information to a structured unit at the financial matter level. Then, market status representations corresponding to the event impact units are generated based on market status data, and long-term interest features and conversation intent features are extracted based on user interaction data to construct a decision-making state representation for the target user. On this basis, combining the event impact units, market status representation, and decision-making state representation, a causal association estimation result including exposure bias and timeliness bias is established, and a candidate event impact unit set is determined accordingly. Finally, distribution calibration and intra-cluster mapping processing under confidence constraints are performed on the candidate event impact unit set to generate personalized financial information recommendation results.

[0017] Through the above technical solutions, this application can achieve unified modeling and deduplication control of similar financial matters in the process of financial information recommendation, thereby reducing the duplicate recommendation rate; at the same time, it can correct exposure bias and timeliness bias, thereby reducing recommendation bias; and by combining market status and user decision-making status for collaborative sorting, it can improve the timeliness adaptability and personalization accuracy of recommendation results.

[0018] The technical solution of this application will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0019] Figure 1This is a flowchart illustrating the personalized content recommendation method for financial information provided in this application embodiment. Figure 1 As shown, the method may specifically include: S101: Obtain financial information data, market status data and user interaction data corresponding to the target user; perform semantic parsing, event normalization and cross-source association processing on the financial information data; and generate event impact units corresponding to financial matters. S102, Based on market state data, extract time-series fluctuation characteristics and theme evolution characteristics to generate market state representations corresponding to each event-affected unit; S103, extract long-term interest features and conversation intent features based on user interaction data, and construct a decision state representation corresponding to the target user by combining event impact units; S104. Based on the event impact unit, market state representation, and decision state representation, establish the causal relationship estimation results including exposure bias and timeliness bias, and determine the candidate event impact unit set based on the causal relationship estimation results. S105, based on the set of candidate event impact units, market state representation, and decision state representation, performs distribution calibration sorting and intra-cluster mapping processing under confidence constraints to generate personalized financial information recommendation results for the target user.

[0020] In some embodiments, semantic parsing, event normalization, and cross-source correlation processing are performed on financial information data to generate event impact units corresponding to financial events, including: Perform domain semantic encoding, structured segmentation, and financial element extraction on financial information data to determine the event subject, event object, event attributes, time constraints, and evidence fragments; Based on the financial elements, the system performs event trigger identification, numerical fact alignment, semantic reference resolution, and event boundary verification to generate event fragments corresponding to each financial matter. Based on the entity association degree, semantic consistency degree, temporal proximity degree, numerical matching degree and market response similarity of event fragments, cross-source aggregation, same-source deduplication and normalization mapping are performed, and the aggregation results are checked for consistency according to the source credibility constraint to generate event impact units carrying the impact duration attribute.

[0021] Specifically, the system first accesses financial information data within the target time range, performing unified preprocessing on news flashes, in-depth reports, announcement interpretations, and research report excerpts from different sources to eliminate headline noise, format differences, reprint marks, and the impact of repeated paragraphs. Then, based on the semantic features of the financial domain, it performs domain-specific semantic encoding and structured segmentation on the information content, extracting the main subject, the object of action, attribute changes, time conditions, and evidence content that characterize financial events from the text. Building upon this, it further identifies the triggering semantics, numerical changes, and contextual referential relationships of financial events, forming event fragments corresponding to each financial event. Finally, it performs cross-source aggregation, same-source deduplication, and normalization mapping on event fragments from different information sources, and combines source credibility constraints and market response information to perform consistency verification on the aggregation results, generating event impact units carrying impact duration attributes, providing unified input for subsequent market state modeling, user decision state modeling, and recommendation ranking.

[0022] In this embodiment, domain semantic encoding refers to using a semantic representation model pre-trained on financial corpus and enhanced with an industry lexicon to map information text into vector representations containing entity semantics, event relationship semantics, and numerical change semantics; structured segmentation refers to decomposing information text into layers according to title, lead, main body sentence, numerical description paragraph, and quotation paragraph in order to separate key facts from background narratives.

[0023] Financial element extraction refers to extracting the event subject, event object, event attributes, time constraints, and evidence fragments from the segmentation results. Among them, the event subject can correspond to listed companies, industry sectors, financial products, or macroeconomic objects; the event object can correspond to affected objects such as profits, revenue, ratings, shareholding ratios, and issuance scales; event attributes are used to describe the types of changes such as upward adjustment, downward adjustment, growth, acquisition, reduction of holdings, and repurchase; time constraints are used to limit the time of announcement, effective time, or statistical period; and evidence fragments are used to preserve the original text content supporting the financial event.

[0024] Event boundary verification refers to determining the scope of the context before and after an event is triggered, avoiding the misinclusion of background information, historical reviews, and commentary into the target event. Impact duration attribute refers to assigning duration markers such as short-term disturbance, phased duration, or medium-to-long-term effect to the event's impact unit based on the event type, factual intensity, time conditions, and subsequent market response characteristics.

[0025] For example, on the morning of a trading day, the system received three pieces of financial news from different sources. The first piece of news was titled "Company A releases annual performance forecast, expecting net profit to increase by 35% to 45% year-on-year"; the second piece of news was that Company A expects a significant improvement in profitability this year, with net profit growth ranging from 35% to 45%; the third piece of news was that, influenced by the news of the expected profit increase, Company A's stock price rose rapidly after the market opened, and the consumer electronics sector to which it belongs also became active.

[0026] The system first performs domain semantic encoding and structured segmentation on the three pieces of information. From the first two pieces of information, it identifies the subject of the event as Company A, the object of the event as net profit, the event attribute as expected profit increase, the time constraint as the disclosure time of the annual performance forecast, and the evidence fragment as the original text content such as expected net profit increase of 35% to 45%. For the third piece of information, it identifies the subject of the event as Company A and the consumer electronics sector, the object of the event as stock price and sector activity, and the event attribute as enhanced market response.

[0027] Subsequently, the system performs event trigger identification based on financial elements, categorizes semantics such as the release of earnings forecasts, profit improvement, and expected increase into the same earnings expected increase trigger type, confirms consistency in the 35% to 45% increase range through numerical fact alignment, confirms that Company A and the Company refer to the same entity through semantic reference resolution, and eliminates generalized descriptions of industry cycles in analyst comments through event boundary verification, thus forming two event fragments: the Company A earnings expected increase fragment and the Company A earnings expected increase triggering market response fragment.

[0028] Furthermore, the system calculates the entity correlation, semantic consistency, temporal proximity, numerical matching, and market response similarity among the event fragments. For the earnings growth fragments formed by the first two news items, since the main entities are consistent, the event semantics are consistent, the release time is close, and the numerical ranges are completely matched, the system aggregates them across sources into the same event and performs same-source deduplication on duplicate content in the reprinted sources; for the market response fragment formed by the third news item, the system identifies that there is an event-driven relationship between it and the earnings growth fragment, and therefore includes it as a supplementary response fragment for that event in the aggregation result.

[0029] Meanwhile, the system performs consistency checks on the aggregation results based on source credibility constraints. For example, if one piece of information comes from a highly credible announcement interpretation source and the other comes from a regular information reprint source, the numerical facts and time constraints in the former will be retained first. If there is a conflict between different sources regarding the increase range, the original announcement parsing result will be used as the main fact, and the conflict description will be marked as information to be downweighted.

[0030] Subsequently, the system retrieved the market changes within a preset observation window after the disclosure of the event and found that Company A's stock price rose and trading activity increased, and the sector linkage was strengthened. Based on this, it was determined that in addition to the announcement trigger attribute, the event also had the characteristics of phased market dissemination, and its impact duration attribute was assigned as phased continuous type.

[0031] Ultimately, the system generates an event impact unit corresponding to the expected increase in Company A's performance. This event impact unit uniformly encapsulates the main entity, event attributes, numerical facts, time constraints, evidence fragments, reliable source results, and impact period attributes.

[0032] Through the above implementation methods, the same financial matters that were originally scattered across different information sources can be uniformly identified, merged, and structurally represented. By using market response information and source credibility constraints, the consistency and usability of the extracted results can be improved, thereby providing a stable data foundation for event-level deduplication, timeliness modeling, and user interest matching in subsequent recommendation processing.

[0033] In some embodiments, time-series fluctuation characteristics and thematic evolution characteristics are extracted based on market state data to generate market state representations corresponding to each event-affected unit, including: Perform time anchoring, window segmentation, and multi-scale state alignment on market state data corresponding to event impact units to generate market sequence fragments associated with each event impact unit; Based on market sequence fragments, information on fluctuation strength transformation, trading activity migration, and thematic correlation diffusion is extracted to generate time-series fluctuation characteristics and thematic evolution characteristics; Joint encoding, attention constraint mapping, and state consistency verification are performed on the time-series fluctuation characteristics and thematic evolution characteristics to generate market state representations corresponding to each event-affected unit.

[0034] Specifically, after constructing the event impact units, the system further retrieves market state data associated with each event impact unit. Using the event occurrence time, impact duration attributes, main entities, and related themes encapsulated in the event impact units as constraints, the system performs time anchoring, window segmentation, and multi-scale state alignment on the market state data to form market sequence segments corresponding to each event impact unit. Then, based on each market sequence segment, the system extracts time-series fluctuation features and theme evolution features reflecting the market state change process. On this basis, the system performs joint encoding, attention constraint mapping, and state consistency verification on the time-series fluctuation features and theme evolution features to generate market state representations corresponding to each event impact unit, which can then be used for subsequent candidate event impact unit screening, causal correlation estimation, and calibration ranking processing.

[0035] In this embodiment, market status data refers to a set of data related to the financial information recommendation process that characterizes the market's operational status, including price change data, transaction change data, sector linkage data, theme popularity change data, and capital activity change data. Time anchoring refers to determining the alignment starting point of market status data based on the event occurrence time, disclosure time, or market response start time corresponding to the event impact unit; window segmentation refers to constructing a pre-observation window, an immediate response window, and a subsequent propagation window around the alignment starting point to distinguish the pre-event background state, the state at the time of event triggering, and the state after event diffusion; multi-scale state alignment refers to mapping minute-level, hourly-level, and daily-level market status data to a unified analysis coordinate system to maintain the correlation between data at different time scales.

[0036] The volatility intensity conversion information is used to characterize the process of market changes from a stable state to an active state, and from an active state to a declining state; the trading activity migration information is used to characterize the process of trading attention being transferred between individual stocks, sectors and related themes; the theme correlation diffusion information is used to characterize the diffusion intensity, diffusion direction and diffusion level of the market theme corresponding to the event impact unit within a preset time range.

[0037] In this embodiment, the transaction activity migration information is not only used to characterize changes in the transaction activity of the event subject itself, but also to characterize the process by which transaction attention spreads from the event subject as the starting point, along entity relationships, sector affiliation relationships, and theme relationships, to other market objects. Specifically, the system takes the main entity in the event impact unit as the starting node, establishes a market relationship path based on the industry sector to which the main entity belongs, related industry chain objects, and theme tags, and calculates the changes in transaction activity of the main entity, related stocks, sector, and related theme in different time windows. If the transaction volume of the main entity increases first, and subsequently multiple related objects under the same sector or theme show increases in transaction volume, turnover rate, or transaction attention, the system identifies this continuous change as a transaction activity migration process and generates transaction activity migration information. This transaction activity migration information includes the starting active object, migration direction, migration time difference, response strength of related objects, and theme layer diffusion strength, which are used to characterize the state of transaction attention being transmitted from the individual stock layer to the sector layer and theme layer.

[0038] For example, in some examples, after generating market sequence fragments, the system performs feature extraction processing on market data in the pre-observation window, the immediate response window, and the subsequent propagation window. Specifically, within the immediate response window, the system detects that Company A's price increase has shifted from low-level fluctuations to a rapid upward trend, and its trading volume has increased from a normal level to 2.6 times that of the same period of the previous trading day, thereby extracting information on the shift in volatility from weak to strong fluctuations. At the same time, the system does not directly treat Company A's own increase in trading volume as complete information on the migration of trading activity. Instead, it further establishes a trading attention propagation path starting from Company A, along with the consumer electronics sector to which Company A belongs, the performance growth theme, and related entities in the industry chain, continuously detecting changes in the trading activity of related market entities in the subsequent window.

[0039] For example, after the system detected a significant increase in the trading volume of Company A between 09:30 and 10:30 on April 15, 2026, several entities within the consumer electronics sector that have business or thematic connections with Company A subsequently experienced increased trading volume and turnover rates between 10:00 and 11:30. Specifically, the overall trading volume of the sector increased by 1.8 times compared to the same period of the previous trading day, and the click-through rate and search volume for information related to the performance growth theme continued to rise in the afternoon of that day. Based on this, the system determined that the trading attention was first triggered at the individual stock level of Company A, then spread to the consumer electronics sector level, and further transmitted to the performance growth theme level, thus generating trading activity migration information consisting of initial activity in individual stocks, synchronous response from the sector, and continuous diffusion of the theme. This trading activity migration information, together with volatility strength conversion information, is used to form time-series volatility characteristics and participates in the joint coding of subsequent market state representation along with the theme-related diffusion information.

[0040] Furthermore, within the subsequent dissemination window, the system detected a simultaneous increase in trading activity across multiple stocks within the consumer electronics sector, along with a continuous rise in the frequency of earnings growth-themed news releases, user searches, and browsing frequency of related objects. This allowed the system to extract thematic diffusion information related to the expected earnings increase, which then spread to the sector and theme levels. The system uses the increase in Company A's own trading volume as the starting point of the trading activity migration, the enhanced trading activity of related objects within the consumer electronics sector as the intermediate state, and the continuous increase in attention to the earnings growth theme as the extended state. This forms trading activity migration information that characterizes the process of cross-object, cross-sector, and cross-theme transmission of trading attention.

[0041] Joint coding refers to mapping temporal fluctuation characteristics and thematic evolution characteristics into a unified state coding space to form a joint market representation; attention constraint mapping refers to assigning higher attention weights to key time areas and key thematic areas in the joint market representation based on the main entities, event attributes, impact duration attributes, and thematic constraints in the event impact unit; state consistency verification refers to comparing the consistency of representation results formed under different windows, different scales, and different thematic paths to eliminate abnormal responses caused by occasional noise.

[0042] For example, the system has generated an event impact unit corresponding to Company A's expected profit increase based on financial information data. This event impact unit includes Company A as the event subject, the event attribute as expected profit increase, the event occurrence time as 09:20 on April 15, 2026, the impact duration attribute as phased and continuous, and the related themes as consumer electronics and profit growth. Based on this, the system retrieves market status data for Company A, the consumer electronics sector, and related thematic indices from April 14, 2026 to April 18, 2026, and uses 09:20 on April 15, 2026 as the time anchor point for window segmentation. Specifically, market data from the trading day before the event is determined as the pre-observation window, market data from the 120 minutes after the event is determined as the immediate response window, and market data from the three trading days after the event is determined as the subsequent dissemination window. Subsequently, the system further performs multi-scale state alignment on minute-level trading volume change sequences, hourly-level sector rotation sequences, and daily-level thematic popularity change sequences to generate market sequence fragments associated with this event impact unit.

[0043] Furthermore, after generating market sequence fragments, the system performs feature extraction processing on market data in the pre-observation window, the immediate response window, and the subsequent propagation window. Specifically, within the immediate response window, the system detects that Company A's price increase has shifted from low-level fluctuations to a rapid upward trend, and the trading volume has increased from the normal level to 2.6 times that of the same period of the previous trading day. From this, it extracts information on the shift from weak to strong fluctuations and information on the migration of trading activity from low to high activity. Within the subsequent propagation window, the system detects that the trading activity of multiple stocks in the consumer electronics sector has increased simultaneously, and the frequency of the release of earnings growth-themed information and the frequency of user attention have continued to rise. From this, it extracts information on thematic diffusion surrounding the earnings growth announcement, which spreads to the sector and theme levels.

[0044] Then, the system inputs the extracted temporal fluctuation features and theme evolution features into a unified state coding space for joint coding to form a joint market representation. Based on the main entities, event attributes, and impact duration attributes in the event impact unit, higher attention weights are applied to the high response time area from 30 minutes after the event to the closing stage, as well as the highly correlated theme areas corresponding to consumer electronics and performance growth, to generate target state mapping results.

[0045] The system then compares the consistency between the minute-level, hourly-level, and daily-level state mapping results. If a short-term spike appears only in a few minutes and does not continue in the hourly and daily paths, it is identified as an abnormal disturbance and its weight is adjusted. If sector activity and theme diffusion remain continuous across multiple windows, the response is confirmed to have sustained propagation characteristics. Finally, the system generates a market state representation corresponding to the impact unit of Company A's expected earnings increase event. This representation comprehensively reflects the degree of market activity improvement, sector linkage, and theme diffusion after the event is triggered.

[0046] Through the above implementation methods, market state data can be oriented and hierarchically modeled around event-influenced units. This allows market state representation to move beyond simply referencing single price or transaction indicators. Instead, it can depict the state change process, theme diffusion path, and cross-scale response relationship after an event is triggered. This improves the correlation and consistency between market state representation and event-influenced units, providing a more accurate basis for market state in the subsequent recommendation process, including screening, timeliness judgment, and ranking modeling of candidate event-influenced units.

[0047] In some embodiments, joint encoding, attention constraint mapping, and state consistency verification are performed on time-series fluctuation characteristics and thematic evolution characteristics to generate market state representations corresponding to each event-affected unit, including: By inputting temporal fluctuation features and thematic evolution features into a unified state coding space, cross-scale correlation modeling, temporal dependency propagation, and feature coupling alignment are performed to generate a joint market representation. Based on the event semantic constraints, time constraints, and impact duration constraints corresponding to the event impact unit, attention weight allocation and response region mapping are performed on the joint market representation to generate target state mapping results. Based on the target state mapping results, cross-window consistency comparison, abnormal response verification, and state association strength correction are performed, and the verified target state mapping results are determined as the market state representation corresponding to the event impact unit.

[0048] Specifically, the system first inputs the temporal fluctuation characteristics and thematic evolution characteristics from different time scales, different market objects and different thematic paths into a unified state coding space. Within the unified state coding space, it completes cross-scale correlation modeling, temporal dependency propagation and feature coupling alignment to form a joint market representation that can reflect the overall market response relationship after an event is triggered.

[0049] Then, the system combines the event semantic constraints, time constraints, and impact duration constraints encapsulated in the event impact unit to perform attention weight allocation and response region mapping on the key time region, key object region, and key theme region in the market joint representation, generating the target state mapping result; finally, the system performs cross-window consistency comparison, abnormal response verification, and state association strength correction on the target state mapping result, and determines the target state mapping result that has passed the verification and been corrected as the market state representation corresponding to the event impact unit.

[0050] In this embodiment, the unified state coding space refers to a common expression space used to accommodate features from different time domains, object granularities, and thematic paths. This space can map heterogeneous features such as minute-level fluctuations, hourly-level linkages, and daily-level evolutions to a unified dimension for joint analysis. Cross-scale correlation modeling refers to establishing the correspondence between minute-level price and transaction changes, hourly-level sector rotation, and daily-level thematic diffusion, so that market states are no longer limited to single-time-domain observation results.

[0051] Temporal dependency propagation refers to the propagation of early local responses to subsequent plate responses and thematic responses according to the state transmission order after an event is triggered, forming a continuous state dependency chain. Feature coupling alignment refers to mapping the active switching and strength changes in temporal fluctuation features to the heat diffusion and correlation migration in thematic evolution features, eliminating the representational bias caused by different observation scales.

[0052] Semantic constraints refer to the constraint information derived from the event's impact unit, such as the event subject, event attributes, related themes, and affected objects; temporal constraints refer to the event occurrence time, information disclosure time, and the range of the preset analysis window; and impact duration constraints refer to duration attributes such as short-term disturbance type, phase-based continuous type, or medium-to-long-term effect type.

[0053] Response region mapping refers to projecting the time periods, thematic clusters, and object clusters related to the current event's impact unit in the market joint representation as key analysis regions. Anomaly response verification involves identifying anomalous peaks or noise disturbances that occur only within a local short-term window but cannot be continuously verified in adjacent windows or scales. State correlation strength correction refers to enhancing or weakening the correlation between each response region and the event's impact unit based on consistency comparison results.

[0054] For example, regarding the impact of the aforementioned Company A's expected profit increase event, the system has extracted three sets of core features from market state data. These include time-series fluctuation features such as price fluctuations, trading intensity changes, and fluctuation level switching information for Company A within 30 minutes, 120 minutes, and the following three trading days after the event disclosure; and theme evolution features such as enhanced correlation information within the consumer electronics sector on an hourly scale and the diffusion of the profit growth theme's popularity on a daily scale. The system first inputs these features into a unified state coding space and establishes cross-scale correlation paths between minute-level individual stock changes, hourly sector changes, and daily theme changes.

[0055] For example, the system detects that after Company A disclosed its expected profit increase at 09:20 on April 15, 2026, its trading volume increased to 2.6 times that of the same period of the previous trading day between 09:30 and 10:30, and the price increase rapidly amplified. At the same time, the consumer electronics sector saw a coordinated rise after 11:00, and information related to profit growth continued to increase after 14:00 that day. Based on this, the system performs time-dependent propagation and feature coupling alignment of the strong volatility switching at the individual stock level, the enhanced linkage at the sector level, and the diffusion of popularity at the theme level, generating a joint market representation.

[0056] Subsequently, the system retrieved the constraint information from the event's impact unit. The semantic constraints indicated that the event was a pre-earnings increase event, with Company A as the subject and related themes of consumer electronics and earnings growth. The time constraint indicated that the event was disclosed at 09:20 on April 15, 2026. The impact duration constraint indicated that the event was a phase-based event. Based on these constraints, the system assigned a higher attention weight to the period from 09:20 on April 15, 2026 to the market close in the market joint representation, and focused on mapping the response areas corresponding to the consumer electronics sector and the earnings growth theme, while reducing the attention weight for short-term fluctuation areas of other sectors unrelated to the event. After the response area mapping, the system obtained the target state mapping result, which reflected the market transmission path of Company A's pre-earnings increase event from the initial response of the individual stock to the diffusion to the sector and then to the strengthening of the theme.

[0057] Furthermore, after obtaining the target state mapping result, the system performs a cross-window consistency comparison. Specifically, the system compares the high-weighted region in the immediate response window with the corresponding region in the subsequent propagation window. If a price spike is found to appear only in the 5-minute interval from 09:35 to 09:40 and not continue in the subsequent 30-minute window, hourly sector path, and daily theme path, the spike is identified as an abnormal response and is downweighted. If the increased activity of Company A's stock, the increased linkage of the consumer electronics sector, and the rising popularity of the performance growth theme are found to be continuous across multiple windows and scales, the mapping result is determined to have high consistency.

[0058] Subsequently, the system corrects the state correlation strength of relevant regions based on the consistency comparison results. For example, the correlation strength of the theme diffusion path that remained active for three consecutive trading days was increased from 0.62 to 0.81, while the correlation strength of the local abnormal path that appeared only in a single period was decreased from 0.48 to 0.19. Finally, the system determines the target state mapping result that has passed the verification and has been corrected for correlation strength as the market state representation corresponding to the impact unit of Company A's expected earnings increase event. This representation reflects both the immediate impact of the event on the individual stock level and its stage propagation characteristics on the sector and theme levels.

[0059] In this embodiment, association strength refers to the degree of matching between different technical objects in semantic, temporal, thematic, market state, or user state dimensions, used to characterize whether there is a basis for association between them. The value of association strength ranges from 0 to 1. The closer the value is to 1, the higher the degree of matching between the two in multiple dimensions; the closer the value is to 0, the less stable the association between the two. Association strength can be used to represent the matching relationship between event-influencing units and user decision-making state representations, or it can be used to represent the matching relationship between event-influencing units and market state representations. For example, the event subject, event attributes, impact period attributes, and associated themes of the event-influencing unit "Company A's performance is expected to increase" are matched with the objects of interest, preferred event types, timeliness preferences, and theme preferences in the target user's current decision-making state representation, respectively. After normalization and weighting, the association strength is 0.86. This value indicates that the event-influencing unit has a high degree of matching with the target user's current decision-making state, rather than indicating the certainty of click probability or recommendation results.

[0060] Through the above implementation methods, market response features from different time domains and levels can be collaboratively modeled in a unified state coding space. Guided by event semantic constraints, time constraints, and impact duration constraints, directional mapping can be achieved. Furthermore, cross-window consistency comparison and abnormal response verification can improve the stability and credibility of the representation results. This enables the generated market state representation to more accurately reflect the real market response relationship corresponding to the event impact unit, providing a more refined market state basis for candidate screening, bias estimation, and ranking decisions in subsequent recommendation processing.

[0061] In some embodiments, long-term interest features and session intent features are extracted based on user interaction data, and a decision state representation corresponding to the target user is constructed by combining event influence units, including: The user interaction data is stratified by time, the behavior sequence is normalized, and the preference intensity is quantified to generate historical interaction trajectories corresponding to the target user. Based on historical interaction trajectories, long-term temporal memory encoding, interest topic aggregation, and stable preference constraints are performed to generate long-term interest features. Based on interaction fragments within the current session, context dependency modeling, intent focus identification, and short-term drift verification are performed to generate session intent features. Based on the correlation and response relationships between long-term interest features, session intent features, and event-influencing units, user event collaborative alignment, decision tendency mapping, and state consistency correction are performed to generate a decision state representation corresponding to the target user.

[0062] Specifically, the system first acquires user interaction data of the target user on the financial information platform, including exposure, clicks, dwell time, favorites, sharing, searches, subscriptions, and ignores. It then performs time-level stratification, behavior sequence regularization, and preference intensity quantification on the user interaction data according to a preset time granularity, generating a historical interaction trajectory corresponding to the target user. Next, based on the historical interaction trajectory, it performs long-term temporal memory encoding, interest topic aggregation, and stable preference constraints to generate long-term interest features. Simultaneously, based on interaction fragments within the current session, it performs context-dependent modeling, intent focus identification, and short-term drift verification to generate session intent features. Furthermore, the system combines the aforementioned generated event impact units with the correlation response relationships between long-term interest features, session intent features, and event impact units to perform user event co-alignment, decision tendency mapping, and state consistency correction, generating a decision state representation corresponding to the target user for subsequent causal correlation estimation, candidate event impact unit screening, and calibration ranking.

[0063] It should be noted that all actions requiring the acquisition of the target user's privacy data in this application embodiment require the user's authorization. For example, a privacy data acquisition agreement can be sent to the user, and the corresponding privacy data can only be acquired after obtaining the user's confirmation and authorization.

[0064] In this embodiment, time stratification refers to dividing user interaction data into a long-term observation layer, a recent activity layer, and a current session layer to distinguish between stable preferences and immediate attention; behavior sequence normalization refers to mapping multiple types of behaviors corresponding to the same user into a unified time series according to the order of interaction, and retaining information such as behavior object, occurrence time, behavior type, interaction intensity, and related topics; preference intensity quantification refers to assigning differentiated weights to different interactions based on click depth, dwell time, frequency of secondary visits, collection behavior, and active search behavior.

[0065] Long-term temporal memory encoding refers to extracting thematic preferences, object preferences, and attention style preferences that have persisted over a long period of time based on historical interaction trajectories within a long-term observation layer; interest theme aggregation refers to merging multiple interactions related to the same financial theme, the same industry chain, or the same investment clue into a unified interest cluster; stable preference constraint refers to assigning higher stability labels to consistent themes that maintain a high response over multiple time periods.

[0066] Context-dependent modeling refers to identifying the coherent direction of a user's current browsing behavior by focusing on the thematic, object, and event connections between adjacent interactions in the current session; intent focus identification refers to determining the user's current focus, the type of matter, and the direction of decision-making from the current session interaction fragments; short-term drift verification refers to detecting whether there is a significant deviation between the content of the current session and long-term stable preferences.

[0067] User event collaborative alignment refers to mapping the long-term interest characteristics and session intent characteristics formed on the user side to the event subject, event attributes, impact duration attributes, and related topics in the event impact unit; decision tendency mapping refers to mapping the user's response tendency to different event types, different duration attributes, and different topic directions to a calculable state intensity; state consistency correction refers to checking the response differences between long-term interest characteristics and session intent characteristics on the same event impact unit to eliminate the offset caused by occasional clicks or abnormal jumps.

[0068] For example, user interaction data for target user B over the past 30 days shows that they repeatedly clicked on information related to semiconductors, artificial intelligence computing power, consumer electronics, and earnings forecasts. They also spent considerable time on multiple articles discussing company performance, industry chain changes, and sector-wide analyses. Specifically, the average time spent on information related to Company A, chip company B, and related consumer electronics sectors exceeded 95 seconds, and they made four bookmarks and three active searches. The system first stratifies the user interaction data for the past 30 days, the past 7 days, and the current session by time, and then organizes continuous interactions such as clicks after exposure, pauses after clicks, and bookmarks after pauses into historical interaction trajectories.

[0069] Subsequently, the system performed long-term temporal memory encoding based on the historical interaction trajectory of the past 30 days, identifying that the user had been continuously paying attention to themes such as performance growth, the semiconductor chain, and the recovery of the consumer electronics market over a long period of time, thus forming long-term interest characteristics. At the same time, the system performed context-dependent modeling on the most recent 12 interaction segments in the current session, and found that the user continuously browsed content such as Company A's performance increase forecast, abnormal movements in the consumer electronics sector, and the spread of the performance growth theme between 09:10 and 10:20 on April 15, 2026. After browsing the relevant information of Company A, the user continued to search for leading companies in the consumer electronics sector with performance increase forecasts. Thus, the system identified that the intent focus of the current session was to find consumer electronics theme opportunities with phased dissemination characteristics around performance increase forecasts, and generated corresponding session intent characteristics.

[0070] Subsequently, the system co-aligned long-term interest characteristics and session intent characteristics with the aforementioned event impact unit related to Company A's expected profit increase. Since the event subject of this impact unit is Company A, the event attribute is expected profit increase, the impact period attribute is phased and continuous, and the associated themes are consumer electronics and profit growth, it maintains a high degree of consistency with the user's long-term stable themes and the current session's immediate focus. Therefore, the system increased the user's association response strength for this event impact unit to 0.86. Simultaneously, the system also detected that the user briefly clicked on a gold hedging-related news item in the current session, but the dwell time was only 7 seconds, and there was no subsequent continuous interaction on the same topic. Therefore, in the short-term drift verification, this behavior was identified as an occasional shift and not included in the high-intensity decision tendency mapping.

[0071] Furthermore, the system integrates the user's long-term preferences and conversational intent regarding the duration of the impact of events with expected earnings growth and the consumer electronics theme, generating corresponding intermediate results of decision-making tendencies. Through state consistency correction, the system eliminates the impact of occasional biases, ultimately obtaining the decision-making state representation for target user B. This decision-making state representation indicates that user B currently prefers to receive financial information related to events with expected earnings growth, the spread of the consumer electronics theme, and the stage of its propagation, while maintaining a lower response weight for short-term noise information unrelated to the current conversation direction.

[0072] Through the above implementation methods, users' long-term stable interests and immediate decision-making concerns in the current session can be hierarchically modeled and collaboratively integrated. Under the constraints of event influence units, user event collaborative alignment and state correction can be completed. This allows the generated decision state representation to reflect both the user's stable financial focus over a long period and the short-term focus in the current session, thereby improving the personalized adaptation capability, timeliness matching capability, and the ability to depict the user's true decision-making intent in the subsequent candidate event influence unit screening and recommendation ranking process.

[0073] In this embodiment, response strength refers to the comprehensive quantitative value of the effective interactive response generated by the target user to the event-influencing unit under the combined effects of the target user, the event-influencing unit, and the market state representation. The response strength ranges from 0 to 1. The closer the value is to 1, the more likely the event-influencing unit is to trigger effective interactive responses such as reading, prolonged stay, collection, searching, or continuous browsing under the current market state and user state; the closer the value is to 0, the lower the probability of an effective interactive response. Response strength can be calculated by fusing attention strength, association strength, market state matching degree, timeliness response weight, and historical effective interaction evidence. To avoid terminology confusion, the specification will use "attention strength" when referring to the quantification of user preferences, "association strength" when referring to the matching degree between objects, and "response strength" when referring to the prediction of comprehensive interactive responses.

[0074] In some embodiments, based on long-term interest features, session intent features, and the correlation response relationship between event-influencing units, user event collaborative alignment, decision tendency mapping, and state consistency correction are performed to generate a decision state representation corresponding to the target user, including: A multi-temporal preference representation of the target user is established based on long-term interest features and conversation intent features, and the distribution of user attention response corresponding to the event impact unit is determined according to the correlation response relationship; Based on the distribution of user attention responses, semantic coupling alignment, time-sensitive mapping, and preference weight allocation are performed between users and event-affected units to generate intermediate results of decision-making tendencies. Based on the intermediate results of decision tendency, cross-time domain consistency verification, abnormal offset correction and state fusion update are performed to generate the decision state representation corresponding to the target user.

[0075] Specifically, the system first uniformly represents the long-term interest characteristics and conversation intent characteristics of the target user, forming a multi-temporal preference representation that covers long-term stable preferences and current conversation focus; then, combining the information such as event subject, event attributes, impact duration attributes, related topics and evidence strength encapsulated in the event impact unit, the system determines the distribution of user attention response corresponding to each event impact unit based on the correlation response relationship between user-side features and event-side features.

[0076] Based on this, the system performs semantic coupling alignment, time-sensitive mapping, and preference weight allocation between users and event-affected units according to the distribution of user attention responses, generating intermediate results of decision tendency. Finally, it performs cross-time domain consistency verification, abnormal offset correction, and state fusion update on the intermediate results of decision tendency to form a decision state representation corresponding to the target user, providing user-side state input for subsequent candidate event-affected unit screening and distribution calibration sorting.

[0077] In this embodiment, multi-temporal preference representation refers to the comprehensive representation result that maps the stable financial focus direction formed by the user over a long observation period and the immediate focus direction presented in the current session to a unified expression space. Among them, long-term interest features are used to characterize the industry themes, event types and object range that the user continues to pay attention to over a long period of time, and session intent features are used to characterize the local focus of attention formed by the user around a certain matter in the current access phase.

[0078] User attention response distribution refers to the distribution of target users' responses to different event-affected units across the dimensions of topic, object, event type, and timeliness. Semantic coupling alignment refers to determining the degree of coupling between user-side preference semantics and event semantics on the event-affected unit side within a unified semantic space.

[0079] Time-sensitive mapping refers to mapping the duration of an event's impact on the current session's time frame based on differences in user responses to news flashes, intraday market fluctuations, post-mortem analyses, or phased tracking content across different time periods. Preference weighting involves weighting the contributions of each dimension in the user's attention response distribution based on long-term interest characteristics and session intent characteristics.

[0080] Cross-temporal consistency verification refers to comparing whether the response direction and intensity of long-term interest features and session intent features are consistent within the same event-affected unit. Anomaly offset correction refers to identifying localized abnormal attention caused by accidental touches, brief pauses, or occasional jumps, and then downweighting them. State fusion update refers to jointly updating the verified long-term preference response and short-term intent response to generate a representation that reflects the user's current true decision-making and attention status.

[0081] For example, target user C has been consistently following financial news related to semiconductors, consumer electronics, earnings growth forecasts, and institutional research over the past 30 days. Specifically, news related to the consumer electronics supply chain was clicked 28 times, with an average dwell time of 86 seconds, and was saved 5 times and actively searched 4 times. News related to the safe-haven theme of gold was clicked only 3 times, with an average dwell time of 9 seconds, and no saving or secondary search behavior was observed. Based on this user's historical interaction data over the past 30 days, the system extracted long-term interest characteristics, determining that their stable preferences are mainly concentrated on themes such as consumer electronics, earnings growth, and economic recovery.

[0082] Subsequently, in the current session from 09:15 to 10:10 on April 15, 2026, User C continuously browsed three types of information: Company A's expected profit increase, abnormal movements in the consumer electronics sector, and the diffusion path of the profit growth theme. After browsing, User C actively searched for leading companies in the consumer electronics sector with expected profit increases and the theme of sustained performance. This generated the session intent characteristics corresponding to the current session, indicating that the user's current focus was not on generalized reading, but on finding consumer electronics theme content with the potential for sustained dissemination around the expected profit increase.

[0083] The system further maps long-term interest features and session intent features into a unified multi-temporal preference representation and establishes a correlation response relationship with the aforementioned event impact unit of Company A's expected performance increase. Since the event subject in this impact unit is Company A, the event attribute is expected performance increase, the impact period attribute is phased and continuous, and the associated themes are consumer electronics and performance growth, it has a high consistency with user C's long-term preferences and current session intent. Therefore, the system assigns a response strength of 0.84 in the theme dimension, 0.88 in the event type dimension, and 0.81 in the timeliness dimension, thus forming a user attention response distribution for this event impact unit.

[0084] Furthermore, after obtaining the distribution of user attention responses, the system continues to perform semantic coupling alignment, time-sensitive mapping, and preference weight allocation. Specifically, the system identifies that the performance growth semantic in user C's long-term preferences is highly coupled with the performance pre-increase event attribute in the event impact unit, and the phase-based continuous topic opportunity in the session intent is consistent with the event's impact duration attribute, thus increasing the coupling weight of the corresponding dimension; at the same time, considering that user C has a high response to the interpretation of anomalies and subsequent diffusion analysis during the trading session, and that the current time point is in the immediate dissemination stage after the event disclosure, the system further increases the weight of the event impact unit in the time-sensitivity adaptation dimension, generating intermediate results of decision tendency.

[0085] Subsequently, the system performed a cross-time-domain consistency check and found that the response directions of long-term interest features and session intent features on the impact unit of Company A's expected performance increase event were consistent, both pointing to the theme of consumer electronics performance growth. Therefore, the high-weight mapping was retained. On the other hand, the system detected that user C briefly clicked on a piece of short-term abnormal information on international gold prices at 09:42, but the dwell time was only 6 seconds, and there was no subsequent continuous interaction on the same topic. Therefore, this behavior was identified as an abnormal offset, and its attention intensity was reduced from 0.36 to 0.08 during the abnormal offset correction phase.

[0086] In this embodiment, attention intensity refers to the quantified value of user preference formed by a target user for a specific financial topic, event type, or market object within a preset time window. The value of attention intensity ranges from 0 to 1. The closer the value is to 1, the more sufficient the effective interaction of the target user with the corresponding topic, event type, or market object; the closer the value is to 0, the weaker the evidence of the corresponding interaction. The system normalizes behavioral data such as click count, dwell time, repeated browsing, favorites, active searches, sharing, ignoring, and quick exit, and performs weighted fusion according to behavioral effectiveness and time decay rules to obtain attention intensity. For example, if a target user has generated 28 effective clicks on the consumer electronics topic in the past 30 days, with an average dwell time of 86 seconds, favorites 5 times, and active searches 4 times, the system can calculate the attention intensity of the consumer electronics topic as 0.82; if the target user only has 3 short clicks on the gold hedging topic, with an average dwell time of 9 seconds and no subsequent behavior, the corresponding attention intensity can be calculated as 0.18.

[0087] Finally, the system performs state fusion update on the verified long-term response results and the session response results to obtain the decision state representation corresponding to the target user C. This decision state representation clearly indicates that user C currently has a high probability of paying attention to performance-increase-type and phase-continuous event-affected units under the consumer electronics theme, while maintaining a low response level to occasional topics unrelated to the current decision chain.

[0088] Through the above implementation methods, a unified multi-temporal preference representation can be established between long-term interest features and conversation intent features. Based on event impact units, the distribution of user attention responses corresponding to each event impact unit is determined, and semantic coupling alignment and time-sensitive mapping are performed. Then, the stability and authenticity of user state modeling results are improved through cross-temporal consistency verification and abnormal offset correction, so that the generated decision state representation is more in line with the target user's current financial attention direction and decision tendency, providing more accurate user-side basis for candidate event impact unit screening, bias estimation and ranking optimization in subsequent recommendation processing.

[0089] In some embodiments, a causal correlation estimation result including exposure bias and time elapsed bias is established, and a set of candidate event influence units is determined based on the causal correlation estimation result, including: The bias estimation input is generated by aligning the execution time of event impact units, market state representations, decision state representations, and historical exposure interaction records, identifying intervention factors, and normalizing samples. Based on the bias estimation input, a causal dependency structure is constructed between target users, exposure opportunities, event timeliness and interaction response. Counterfactual association estimation, bias weight correction, exposure offset verification and timeliness decay compensation are performed to generate the corresponding causal association estimation results. Based on the causal correlation estimation results, biased gain prediction, recommendability assessment, candidate screening and correlation ranking are performed on the event impact units to determine the set of candidate event impact units corresponding to the target user.

[0090] Specifically, the system first incorporates event impact units, market state representations, decision state representations, and historical exposure interaction records into a unified processing flow, performing time alignment, intervention factor identification, and sample normalization on various types of data to generate bias estimation inputs. Then, based on the bias estimation inputs, it constructs a causal dependency structure between target users, exposure opportunities, event timeliness, and interaction responses, and performs counterfactual association estimation, bias weight correction, exposure offset verification, and timeliness decay compensation accordingly to generate corresponding causal association estimation results. Finally, based on the causal association estimation results, the system performs biased gain prediction, recommendability assessment, candidate screening, and association ranking on event impact units to determine the set of candidate event impact units corresponding to target users, providing input for subsequent distribution calibration and ranking under confidence constraints.

[0091] In this embodiment, time alignment refers to mapping various types of data to the same analytical timeline based on the event occurrence time of the event-affected unit, the state formation time of the market state representation, the effective time of the user state representation of the decision state, and the exposure time and interaction time in historical exposure and interaction records. Intervention factor identification refers to identifying factors from historical exposure and interaction records that may interfere with user interaction responses but do not directly represent the user's true interests, including exposure ranking, display position, display time period, content popularity, and event timeliness.

[0092] Sample normalization refers to transforming historical exposure samples into standardized analytical samples that include event characteristics, market characteristics, user characteristics, exposure characteristics, and interaction results. Causal dependency structure refers to a structured association model used to describe the interaction path between target users, exposure opportunities, event timing, and interaction responses. Counterfactual association estimation refers to estimating the possible interaction responses of users if the event-affected unit were located at different exposure positions or at different exposure times, while keeping the target user, market state, and event-affected unit constant.

[0093] Bias weighting adjustment refers to adjusting sample weights based on the probability of exposure opportunities to reduce the excessive influence of high-exposure samples on the estimation results. Exposure offset verification identifies abnormally high-click samples caused by homepage top displays or trending topic pushes. Time decay compensation refers to applying time decay correction to the response value of event-affected units that have entered the later stages of propagation but are still relevant to the target user's current decision-making state.

[0094] Debiased gain prediction refers to predicting the increase in the response of a target user's true interest after removing the effects of exposure bias and timeliness bias. Recommendability assessment refers to determining whether an event-affected unit is suitable for inclusion in the candidate set by comprehensively considering the strength of causal relationship, event credibility, market state matching degree, and user decision-making state matching degree.

[0095] For example, when target user C enters the financial information platform at 10:20 AM on April 15, 2026, the system has already generated a corresponding decision-making status representation. This representation shows that user C currently has a high level of attention to performance-driven and phase-based event-influenced units under the consumer electronics theme. Meanwhile, the event-influenced units currently available for recommendation by the system include Company A's performance-driven increase, Company B's share buyback plan, short-term fluctuations in international gold prices, and rumors in a certain new energy sector. The system first retrieves the user's historical exposure and interaction records for the past 14 days and finds that content related to short-term fluctuations in international gold prices has been repeatedly placed in the top two positions on the homepage. Although the click-through rate is high, the average dwell time is only 8 seconds. Content related to Company A's performance-driven increase has had fewer exposures in the past 2 hours, but once clicked, the average dwell time reaches 93 seconds, and it is often accompanied by secondary searches and bookmarking. Based on this, the system performs time alignment and sample normalization on event-influenced units, market status representation, decision-making status representation, and historical exposure and interaction records, and identifies homepage top exposure, hot topic tag push, and event release time differences as the main intervention factors.

[0096] Subsequently, the system constructs a causal dependency structure, analyzes the relationship between user C, exposure opportunity, event timeliness, and interactive response, and performs counterfactual association estimation. The estimation results show that if the impact unit of Company A's expected profit increase event is elevated to the same exposure position as the short-term fluctuation in international gold prices, its expected effective interactive response will be significantly higher than the latter; and the high click-through rate of the short-term fluctuation in international gold prices mainly stems from top exposure intervention. The system further performs biased weighting correction and exposure offset verification, correcting the causal association strength of the short-term fluctuation in international gold prices from 0.71 to 0.34, and the causal association strength of Company A's expected profit increase from 0.56 to 0.83.

[0097] Meanwhile, considering that Company A's expected profit increase occurred within two hours of the event's disclosure, still in the early stages of its dissemination, the system applied a lower level of time-related decay compensation. However, for the event-related unit corresponding to Company B's share repurchase plan, since the event had been disclosed for three trading days and market correlation had weakened, the system applied a higher level of time-related decay correction. Ultimately, based on the causal correlation estimation results, the system performed biased gain prediction and recommendability assessment on each event-related unit, identifying the enhanced correlation between Company A's expected profit increase and the consumer electronics sector as a priority candidate event-related unit, downgrading short-term fluctuations in international gold prices to a low-priority candidate, and eliminating rumors in the new energy sector lacking credible evidence, thus forming a set of candidate event-related units corresponding to target user C.

[0098] Through the above implementation method, the interference of exposure order difference and event timeliness difference on interaction results can be identified and corrected at the candidate generation stage. This makes the candidate selection of event impact units no longer dependent on surface click results, but determined based on causal correlation estimation results that are closer to the user's real interest response. This improves the matching degree between the candidate event impact unit set and the target user's current decision state, and provides more reliable debiased input for subsequent sorting processing.

[0099] In some embodiments, based on the set of candidate event impact units, market state representation, and decision state representation, distribution calibration sorting and intra-cluster mapping processing under confidence constraints are performed to generate personalized financial information recommendation results for the target user, including: The ranking input is constructed based on the set of candidate event impact units, market state representation, and decision state representation, and the ranking confidence constraint is determined by combining the historical interaction sufficiency and state stability of the target user. Based on the ranking confidence constraint, the correlation estimation, distribution bias correction, timeliness consistency adjustment and duplicate association suppression are performed on the candidate event influence unit set to generate a calibrated ranking sequence. Based on the calibration sorting sequence, representative content mapping, output structure adaptation, and result encapsulation are performed on the information content clusters corresponding to the candidate event impact unit set to generate personalized financial information recommendation results for the target user.

[0100] Specifically, the system first integrates the candidate event impact unit set, market state representation, and decision state representation to form a ranking input for the current recommendation time, and determines the ranking confidence constraint by combining the historical interaction sufficiency and state stability of the target user. Then, based on the ranking confidence constraint, the system performs relevance estimation, distribution bias correction, timeliness consistency adjustment, and repetitive association suppression on the candidate event impact unit set to generate a calibrated ranking sequence. Finally, based on the calibrated ranking sequence, the system performs representative content mapping, output structure adaptation, and result encapsulation on the information content clusters corresponding to the candidate event impact unit set to generate personalized financial information recommendation results for the target user.

[0101] In this embodiment, the sorting input refers to the set of inputs to be sorted formed by uniformly splicing together the event attributes, impact duration attributes, related topics, credibility information, activity level, diffusion level, and duration in the market status representation of the candidate event impact unit set, as well as the user attention intensity, time sensitivity, and topic preference intensity in the decision status representation.

[0102] Historical interaction sufficiency refers to the sufficiency of valid interaction samples formed by the target user within a preset observation period, which reflects whether the user profile has a stable modeling foundation; state stability refers to the consistency between the target user's current conversation intent and long-term interests, which reflects whether there is a drastic drift in the current decision state.

[0103] Ranking confidence constraints refer to constraints that limit the credibility of each scoring item during the ranking process based on the sufficiency of historical interactions and the stability of the state. Distribution bias correction refers to adjusting the concentration shift of candidate event impact units in terms of topic distribution, event distribution, and time period distribution. Timeliness consistency adjustment refers to increasing or decreasing the ranking weight of candidate event impact units based on the degree of fit between the current recommendation time and the event impact time period attribute. Duplication association suppression refers to compressing event impact units within the same information content cluster that are semantically similar, factually related, or have repeated propagation chains.

[0104] Content mapping refers to selecting the information content within the same information content cluster that best represents the core facts of the event's impact unit, the current market response, and the user's focus. Output structure adaptation refers to adapting the recommendation results to a newsletter-style, summary-style, or in-depth-read output structure based on the target user's reading habits and current access scenario.

[0105] For example, when target user C enters the financial information platform at 10:20 AM on April 15, 2026, the system has already obtained a set of candidate event impact units corresponding to him / her, including four candidate event impact units: Company A's expected profit increase, enhanced linkage in the consumer electronics sector, Company B's share repurchase plan, and short-term fluctuations in international gold prices. At the same time, the system has generated market state representations corresponding to each candidate event impact unit and generated a representation of user C's current decision-making state. This representation shows that user C has a high response intensity to "performance increase-type and phase-continuous event impact units under the consumer electronics theme".

[0106] The system first constructs a ranking input, integrating the event attributes, impact duration attributes, current market diffusion status, and user-side attention intensity of the four candidate event impact units. Subsequently, the system retrieves user C's interaction history over the past 30 days, finding a high number of effective clicks, long dwell times, favorites, and active searches, indicating a high degree of historical interaction sufficiency. Simultaneously, the current session intent and long-term interests remain consistent on the themes of consumer electronics and performance growth, thus demonstrating high state stability. Based on this, the system determines that the current ranking confidence constraint is at a high level, allowing decision-making state representation and market state representation to occupy a high weight in the ranking process.

[0107] In the ranking process, the system first estimates the correlation of each candidate event's impact unit. Company A's projected earnings increase shows high consistency with User C's long-term interests, current conversational intent, and the stage-diffusion characteristics of the market state representation, thus achieving the highest initial correlation score. While the enhanced linkage within the consumer electronics sector is not a single-company event, it is directly related to the current event propagation path, thus receiving the second-highest correlation score. Company B's buyback plan partially overlaps with the user's long-term preferences, but its correlation with the current conversation focus is weak. Short-term fluctuations in international gold prices have a low correlation with User C's current decision-making state.

[0108] Subsequently, the system performed distribution bias correction and found that the first two candidate event influence units both focused on the consumer electronics topic. If the results were directly output according to the degree of relevance, it would be easy to cause the distribution of results to be too concentrated. Therefore, the distribution of topics in the sequence was moderately balanced. However, due to the high order confidence constraint, the system did not excessively weaken the topic events that were highly relevant to the current session.

[0109] Next, based on the timeliness consistency adjustment rules, the system assigned a higher timeliness adaptation weight to Company A's expected profit increase and the enhanced linkage of the consumer electronics sector, because both are in the early stage of dissemination after the event was disclosed; while Company B's repurchase plan was subject to a moderate reduction in weight, because the matter had been disclosed for two trading days and the market linkage had weakened.

[0110] Subsequently, the system identified that "Company A's expected profit increase" was highly similar to a reprinted article within its cluster in terms of factual description and time constraints. Therefore, it performed duplicate association suppression, retaining only the content with higher credibility and more complete market linkage description as the sorting object, and finally generated a calibrated sorting sequence. The sorting results were as follows: Company A's expected profit increase, enhanced linkage of the consumer electronics sector, and Company B's repurchase plan.

[0111] Furthermore, after obtaining the calibrated sorting sequence, the system performs representative content mapping on the information content clusters corresponding to each sorting item. For the sorting item "Company A's expected profit increase," the system compares the factual completeness, timeliness, and user readability of the announcement interpretation, news bulletin, and market commentary in its corresponding information content cluster, and finally selects one interpretation that simultaneously includes the profit increase range, disclosure time, and sector response description as the representative content; for the sorting item "Strengthened linkage in the consumer electronics sector," the system selects one summary that can summarize the sector's diffusion path and the linkage status of related individual stocks as the representative content.

[0112] Subsequently, based on user C's current browsing scenario and preference for quickly obtaining key information, the system adapts the output structure of the sorting results, encapsulating the first result into a key card of "core facts + growth range + sector response", encapsulating the second result into a supplementary card of "topic diffusion summary + related object linkage", and finally organizing the final results into personalized financial information recommendation results suitable for the current scenario.

[0113] Through the above implementation method, based on the set of candidate event influence units, a ranking confidence constraint that is adapted to the current user state can be determined by combining market state representation and decision state representation. Under this constraint, distribution calibration, timeliness adjustment and duplication suppression are completed. Then, the final recommendation result is generated by mapping representative content within the cluster and adapting the output structure, thereby improving the matching degree between the recommendation result and the target user's current decision focus, the rationality of the result distribution, and the adaptability of the content output.

[0114] In some embodiments, based on ranking confidence constraints, correlation estimation, distribution bias correction, timeliness consistency adjustment, and repetitive association suppression are performed on the candidate event influence unit set to generate a calibrated ranking sequence, including: Based on the ranking confidence constraint, combined with the market state representation and decision state representation, the user association strength, state matching degree, timeliness response weight and ranking confidence level of each candidate event influence unit are calculated to generate the initial ranking result. Based on the initial sorting results, target distribution alignment, deviation constraint correction, cross-topic weight balancing, and dynamic sorting boundary adjustment are performed to generate intermediate sorting results. Based on the intermediate sorting results, the effects of duration consistency adjustment, homogeneous association compression, near-event deduplication suppression, and inter-cluster order correction are performed to generate the final calibrated sorting sequence.

[0115] Specifically, the system first calculates the user association strength, state matching degree, timeliness response weight, and ranking confidence level of each candidate event's impact unit based on ranking confidence constraints, combined with market state representation and decision state representation, to generate an initial ranking result. Then, based on the initial ranking result, the system performs target distribution alignment, deviation constraint correction, cross-topic weight balancing, and dynamic ranking boundary adjustment to generate an intermediate ranking result. Finally, based on the intermediate ranking result, the system performs impact period consistency adjustment, homogeneous association compression, similar event deduplication suppression, and inter-cluster order correction to generate a final calibrated ranking sequence, which serves as the ranking basis for subsequent representative content mapping and recommendation result encapsulation.

[0116] In this embodiment, user association strength refers to the degree of association between the candidate event influencing unit and the target user's current decision-making state, focusing on the comprehensive response level of the user's long-term interest characteristics and conversation intent characteristics to the candidate event influencing unit. State matching degree refers to the consistency between the market state representation corresponding to the candidate event influencing unit and the target user's decision-making state representation, reflecting whether the current market dissemination state aligns with the user's current focus.

[0117] Timeliness response weight refers to the strength of timeliness adaptation calculated based on the event's occurrence time, impact duration, and the current recommendation time. Ranking confidence level refers to the ranking credibility index, which quantifies the acceptability of the current ranking result under constraints of historical interaction sufficiency and state stability. Target distribution alignment refers to ensuring that the ranking results are aligned with the target user's currently acceptable distribution of attention in terms of topic distribution, event type distribution, and duration distribution.

[0118] Deviation constraint correction refers to adjusting the ranking offset caused by localized high popularity, excessive theme concentration, or short-term abnormal fluctuations in the initial ranking results. Dynamic ranking boundary adjustment refers to adaptively adjusting the retention range between high-confidence and low-confidence candidates based on ranking confidence constraints. Influence duration consistency adjustment refers to verifying whether the influence duration attribute of the candidate event's influence unit is consistent with the current market propagation stage.

[0119] Homologous association compression refers to merging and compressing multiple candidate event impact units from the same fact source or the same propagation chain. Approximate event deduplication suppression refers to deduplicating and reducing the weight of candidate event impact units that are semantically highly similar and have overlapping factual expressions. Inter-cluster order correction refers to further correcting the ordering relationship between different information content clusters to avoid any one cluster excessively occupying a leading position.

[0120] For example, when target user C enters the platform at 10:20 on April 15, 2026, the system has already obtained 5 candidate event impact units, namely, Company A's expected profit increase, reprint of Company A's expected profit increase news, enhanced linkage of the consumer electronics sector, Company B's share repurchase plan, and short-term fluctuations in international gold prices. At the same time, the system has determined that the user's ranking confidence constraint is at a high level, because its effective interaction samples in the past 30 days have reached 126, and its current session intent and long-term interests maintain a high degree of consistency on the themes of consumer electronics and profit growth.

[0121] The system first performs an initial ranking calculation on the impact units of the aforementioned candidate events by combining market state representation and decision-making state representation. Among them, the user association strength of Company A's expected profit increase is 0.88, the state matching degree is 0.84, the timeliness response weight is 0.81, and the ranking confidence level is 0.86; the corresponding values ​​for the enhanced linkage of the consumer electronics sector are 0.79, 0.82, 0.78, and 0.83, respectively; although the short-term fluctuations in international gold prices have high real-time popularity, its user association strength is only 0.29 and the state matching degree is only 0.24, so its initial ranking is low.

[0122] Subsequently, the system performed distribution calibration on the initial ranking results and found that two of the top three candidates were concentrated in the same propagation chain of the theme of Company A's expected profit increase. In order to avoid excessive concentration in the sequence, the system increased the cross-theme balancing weight of the consumer electronics sector while retaining the priority of highly relevant themes, constrained and corrected the deviation of the reprint of the news of Company A's expected profit increase, and restricted low-confidence candidates to the later part of the sequence according to the high-confidence ranking boundary, thereby generating the intermediate ranking results.

[0123] Subsequently, the system continued to perform consistency adjustment of the impact period, identifying that although Company B's repurchase plan had a certain relevance, its event had entered the later stage of propagation, so its priority was appropriately lowered; at the same time, the system identified that there was a common source association and similar event relationship between Company A's expected profit increase and the reprinted news of Company A's expected profit increase, so it performed common source association compression and deduplication suppression on the reprinted items, retaining only the event impact units with more complete original interpretation and more sufficient evidence.

[0124] Finally, the system performs inter-cluster order correction on the remaining candidate event impact units to obtain the final calibrated sorting sequence. The sorting results are as follows: Company A's expected profit increase, enhanced linkage of the consumer electronics sector, Company B's repurchase plan, and short-term fluctuations in international gold prices.

[0125] Through the above implementation methods, under the ranking confidence constraint, it is possible to simultaneously consider the user's current decision-making focus, market dissemination status, and event timeliness attributes, and to perform targeted ranking, distribution calibration, and duplication compression on the set of candidate event influence units. This reduces the interference of homologous and similar events on the recommendation sequence, improves the rationality of the ranking results in terms of topic distribution, timeliness adaptation, and user attention matching, and provides a stable foundation for generating personalized financial information recommendation results that better meet the current needs of target users.

[0126] The following are system embodiments of this application, which can be used to execute the method embodiments of this application. For details not disclosed in the system embodiments of this application, please refer to the method embodiments of this application.

[0127] Figure 2 This is a schematic diagram of the structure of a personalized content recommendation system for financial information provided in an embodiment of this application. Figure 2 As shown, this personalized content recommendation system for financial information includes: The acquisition module 201 is used to acquire financial information data, market status data and user interaction data corresponding to the target user, perform semantic parsing, event normalization and cross-source association processing on the financial information data, and generate event impact units corresponding to financial matters. Extraction module 202 is used to extract time-series fluctuation characteristics and theme evolution characteristics based on market state data, and generate market state representations corresponding to each event-affected unit; Module 203 is used to extract long-term interest features and conversation intent features based on user interaction data, and to construct a decision state representation of the target user in combination with event impact units. The determination module 204 is used to establish a causal relationship estimation result including exposure bias and timeliness bias based on the event impact unit, market state representation and decision state representation, and determine the candidate event impact unit set based on the causal relationship estimation result. The generation module 205 is used to generate personalized financial information recommendation results for target users by performing distribution calibration sorting and intra-cluster mapping processing under confidence constraints based on the set of candidate event impact units, market state representation and decision state representation.

[0128] In some embodiments, Figure 2The acquisition module 201 performs domain semantic encoding, structured segmentation, and financial element extraction on financial information data to determine the event subject, event object, event attributes, time constraints, and evidence fragments. Based on financial elements, it performs event trigger identification, numerical fact alignment, semantic reference resolution, and event boundary verification to generate event fragments corresponding to each financial matter. Based on the entity correlation, semantic consistency, time proximity, numerical matching, and market response similarity of the event fragments, it performs cross-source aggregation, same-source deduplication, and normalization mapping, and performs consistency verification on the aggregation results according to the source credibility constraint to generate event impact units carrying the impact period attribute.

[0129] In some embodiments, Figure 2 The extraction module 202 performs time anchoring, window segmentation, and multi-scale state alignment on the market state data corresponding to the event impact units, generating market sequence fragments associated with each event impact unit; based on the market sequence fragments, it extracts volatility strength transformation information, trading activity migration information, and theme association diffusion information to generate time-series volatility features and theme evolution features; it performs joint encoding, attention constraint mapping, and state consistency verification on the time-series volatility features and theme evolution features to generate market state representations corresponding to each event impact unit.

[0130] In some embodiments, Figure 2 The extraction module 202 inputs the temporal fluctuation features and thematic evolution features into the unified state coding space, performs cross-scale correlation modeling, temporal dependency propagation and feature coupling alignment to generate a joint market representation; based on the event semantic constraints, time constraints and impact duration constraints corresponding to the event impact unit, it performs attention weight allocation and response region mapping on the joint market representation to generate a target state mapping result; based on the target state mapping result, it performs cross-window consistency comparison, abnormal response verification and state correlation strength correction, and determines the verified target state mapping result as the market state representation corresponding to the event impact unit.

[0131] In some embodiments, Figure 2 The construction module 203 performs time-leveling, behavior sequence regularization, and preference intensity quantification on user interaction data to generate historical interaction trajectories corresponding to the target user. Based on the historical interaction trajectories, it performs long-term temporal memory encoding, interest topic aggregation, and stable preference constraints to generate long-term interest features. Based on the interaction fragments within the current session, it performs context dependency modeling, intent focus identification, and short-term drift verification to generate session intent features. Based on the correlation and response relationships between long-term interest features, session intent features, and event-influenced units, it performs user event collaborative alignment, decision tendency mapping, and state consistency correction to generate a decision state representation corresponding to the target user.

[0132] In some embodiments, Figure 2The construction module 203 establishes a multi-temporal preference representation of the target user based on long-term interest features and conversation intent features, and determines the user attention response distribution corresponding to the event influence unit according to the associated response relationship; based on the user attention response distribution, it performs semantic coupling alignment, time-sensitive mapping and preference weight allocation between users and event influence units to generate intermediate results of decision tendency; based on the intermediate results of decision tendency, it performs cross-temporal consistency verification, abnormal offset correction and state fusion update to generate the decision state representation of the target user.

[0133] In some embodiments, Figure 2 The determination module 204 performs time alignment, intervention factor identification, and sample normalization on the event impact units, market state representation, decision state representation, and historical exposure interaction records to generate bias estimation input. Based on the bias estimation input, it constructs a causal dependency structure between target users, exposure opportunities, event timeliness, and interaction response, and performs counterfactual association estimation, bias weight correction, exposure offset verification, and timeliness decay compensation to generate corresponding causal association estimation results. Based on the causal association estimation results, it performs bias-free gain prediction, recommendability assessment, candidate screening, and association ranking on the event impact units to determine the set of candidate event impact units corresponding to the target users.

[0134] In some embodiments, Figure 2 The generation module 205 constructs a ranking input based on the candidate event impact unit set, market state representation, and decision state representation, and determines the ranking confidence constraints by combining the historical interaction sufficiency and state stability of the target user; based on the ranking confidence constraints, it performs correlation estimation, distribution bias correction, timeliness consistency adjustment, and repetitive association suppression on the candidate event impact unit set to generate a calibrated ranking sequence; based on the calibrated ranking sequence, it performs representative content mapping, output structure adaptation, and result encapsulation on the information content clusters corresponding to the candidate event impact unit set to generate personalized financial information recommendation results for the target user.

[0135] In some embodiments, Figure 2 The generation module 205, based on the ranking confidence constraint and combined with the market state representation and decision state representation, calculates the user association strength, state matching degree, timeliness response weight, and ranking confidence level of each candidate event's impact unit, generating an initial ranking result; based on the initial ranking result, it performs target distribution alignment, deviation constraint correction, cross-topic weight balancing, and dynamic ranking boundary adjustment to generate an intermediate ranking result; based on the intermediate ranking result, it performs influence period consistency adjustment, homogeneous association compression, similar event deduplication suppression, and inter-cluster order correction to generate the final calibrated ranking sequence.

[0136] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0137] Figure 3 This is a schematic diagram of the electronic device 3 provided in an embodiment of this application. Figure 3 As shown, the electronic device 3 of this embodiment includes: a processor 301, a memory 302, and a computer program 303 stored in the memory 302 and executable on the processor 301. When the processor 301 executes the computer program 303, it implements the steps in the various method embodiments described above. Alternatively, when the processor 301 executes the computer program 303, it implements the functions of each module / unit in the various system embodiments described above.

[0138] Electronic device 3 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 3 may include, but is not limited to, processor 301 and memory 302. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or different components.

[0139] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0140] The memory 302 can be an internal storage unit of the electronic device 3, such as a hard disk or memory of the electronic device 3. The memory 302 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 3. The memory 302 can also include both internal and external storage units of the electronic device 3. The memory 302 is used to store computer programs and other programs and data required by the electronic device.

[0141] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0142] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a readable storage medium (e.g., a computer-readable storage medium). Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which may be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable storage medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0143] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for personalized content recommendation for financial information, characterized by, include: Acquire financial information data, market status data, and user interaction data corresponding to the target user; perform semantic parsing, event normalization, and cross-source association processing on the financial information data to generate event impact units corresponding to financial matters. Based on the market state data, time-series fluctuation characteristics and theme evolution characteristics are extracted to generate market state representations corresponding to each event-affected unit; Based on the user interaction data, long-term interest features and conversation intent features are extracted, and combined with the event impact unit to construct a decision state representation corresponding to the target user. Based on the event impact units, the market state representation, and the decision state representation, a causal correlation estimation result including exposure bias and timeliness bias is established, and a set of candidate event impact units is determined according to the causal correlation estimation result. Based on the set of candidate event impact units, the market state representation, and the decision state representation, distribution calibration sorting and intra-cluster mapping processing under confidence constraints are performed to generate personalized financial information recommendation results for the target user. The step of establishing a causal correlation estimation result including exposure bias and time-sensitivity bias, and determining a set of candidate event influence units based on the causal correlation estimation result, includes: The bias estimation input is generated by aligning the execution time of the event impact unit, the market state representation, the decision state representation, and the historical exposure interaction records, identifying intervention factors, and normalizing the samples. Based on the bias estimation input, a causal dependency structure is constructed between target users, exposure opportunities, event timeliness, and interaction response. Counterfactual association estimation, bias weight correction, exposure offset verification, and timeliness decay compensation are performed to generate the corresponding causal association estimation results. Based on the causal correlation estimation results, biased gain prediction, recommendability assessment, candidate screening and correlation ranking are performed on the event impact units to determine the set of candidate event impact units corresponding to the target user.

2. The method according to claim 1, characterized in that, The process of performing semantic parsing, event normalization, and cross-source association processing on the financial information data to generate event impact units corresponding to financial events includes: The financial information data is subjected to domain semantic encoding, structured segmentation, and financial element extraction to determine the event subject, event object, event attributes, time constraints, and evidence fragments. Based on the aforementioned financial elements, event trigger identification, numerical fact alignment, semantic reference resolution, and event boundary verification are performed to generate event fragments corresponding to each financial matter. Based on the entity association degree, semantic consistency degree, temporal proximity degree, numerical matching degree and market response similarity of the event fragments, cross-source aggregation, same-source deduplication and normalization mapping are performed, and the aggregation results are checked for consistency according to the source credibility constraint to generate event impact units carrying impact duration attributes.

3. The method according to claim 1, characterized in that, The step of extracting time-series fluctuation characteristics and thematic evolution characteristics based on the market state data to generate market state representations corresponding to each of the event-affecting units includes: Perform time anchoring, window segmentation, and multi-scale state alignment on the market state data corresponding to the event impact units to generate market sequence segments associated with each event impact unit; Based on the market sequence segments, information on fluctuation strength transformation, trading activity migration, and theme correlation diffusion is extracted to generate time-series fluctuation characteristics and theme evolution characteristics; Joint encoding, attention constraint mapping, and state consistency verification are performed on the time-series fluctuation features and the theme evolution features to generate market state representations corresponding to each event-affected unit.

4. The method according to claim 3, characterized in that, The process of performing joint encoding, attention-constrained mapping, and state consistency verification on the time-series fluctuation characteristics and the theme evolution characteristics to generate market state representations corresponding to each event-affected unit includes: The temporal fluctuation features and the theme evolution features are input into a unified state coding space, and cross-scale correlation modeling, temporal dependency propagation and feature coupling alignment are performed to generate a joint market representation. Based on the event semantic constraints, time constraints, and impact duration constraints corresponding to the event impact unit, attention weight allocation and response region mapping are performed on the market joint representation to generate target state mapping results. Based on the target state mapping result, cross-window consistency comparison, anomaly response verification, and state association strength correction are performed, and the verified target state mapping result is determined as the market state representation corresponding to the event impact unit.

5. The method according to claim 1, characterized in that, The step of extracting long-term interest features and session intent features based on the user interaction data, and constructing a decision state representation corresponding to the target user in conjunction with the event influence unit, includes: The user interaction data is subjected to time stratification, behavior sequence normalization, and preference intensity quantification to generate historical interaction trajectories corresponding to the target user. Based on the historical interaction trajectory, long-term temporal memory encoding, interest topic aggregation, and stable preference constraints are performed to generate long-term interest features. Based on the interaction fragments within the current session scope, context dependency modeling, intent focus identification, and short-term drift verification are performed to generate session intent features. Based on the correlation and response relationship between the long-term interest features, the session intent features, and the event influence units, user event collaborative alignment, decision tendency mapping, and state consistency correction are performed to generate a decision state representation corresponding to the target user.

6. The method according to claim 5, characterized in that, Based on the correlation and response relationship between the long-term interest features, the session intent features, and the event influence units, the process performs user event collaborative alignment, decision tendency mapping, and state consistency correction to generate a decision state representation corresponding to the target user, including: Based on the long-term interest features and the session intent features, a multi-temporal preference representation of the target user is established, and the user attention response distribution corresponding to the event impact unit is determined according to the associated response relationship; Based on the user attention response distribution, semantic coupling alignment, time-sensitive mapping, and preference weight allocation are performed between users and event-influencing units to generate intermediate decision-making tendencies. Based on the intermediate results of the decision tendency, cross-time domain consistency verification, abnormal offset correction and state fusion update are performed to generate the decision state representation corresponding to the target user.

7. The method according to claim 1, characterized in that, Based on the set of candidate event impact units, the market state representation, and the decision state representation, the process performs distribution calibration sorting and intra-cluster mapping under confidence constraints to generate personalized financial information recommendation results for the target user, including: The ranking input is constructed based on the set of candidate event impact units, the market state representation, and the decision state representation, and the ranking confidence constraint is determined by combining the historical interaction sufficiency and state stability of the target user. Based on the ranking confidence constraints, correlation estimation, distribution bias correction, timeliness consistency adjustment, and duplicate association suppression are performed on the candidate event influence unit set to generate a calibrated ranking sequence. Based on the calibration sorting sequence, representative content mapping, output structure adaptation, and result encapsulation are performed on the information content clusters corresponding to the candidate event impact unit set to generate personalized financial information recommendation results for the target user.

8. The method according to claim 7, characterized in that, The process of performing correlation estimation, distribution bias correction, timeliness consistency adjustment, and duplicate association suppression on the candidate event influence unit set based on the ranking confidence constraint to generate a calibrated ranking sequence includes: Based on the ranking confidence constraint, combined with the market state representation and the decision state representation, the user association strength, state matching degree, timeliness response weight and ranking confidence level of each candidate event influence unit are calculated to generate the initial ranking result. Based on the initial sorting results, target distribution alignment, deviation constraint correction, cross-topic weight balancing, and dynamic sorting boundary adjustment are performed to generate intermediate sorting results. Based on the intermediate sorting results, the following processes are performed: impact period consistency adjustment, homogeneous association compression, near event deduplication suppression, and inter-cluster order correction, to generate the final calibrated sorting sequence.

9. A personalized content recommendation system for financial information, characterized in that, include: The acquisition module is used to acquire financial information data, market status data and user interaction data corresponding to the target user, perform semantic parsing, event normalization and cross-source association processing on the financial information data, and generate event impact units corresponding to financial matters. The extraction module is used to extract time-series fluctuation features and theme evolution features based on the market state data, and generate market state representations corresponding to each of the event impact units. The construction module is used to extract long-term interest features and session intent features based on the user interaction data, and to construct a decision state representation corresponding to the target user in combination with the event influence unit. The determination module is used to establish a causal relationship estimation result including exposure bias and timeliness bias based on the event impact unit, the market state representation and the decision state representation, and to determine a set of candidate event impact units based on the causal relationship estimation result. The generation module is used to perform distribution calibration sorting and intra-cluster mapping processing under confidence constraints based on the set of candidate event impact units, the market state representation, and the decision state representation, to generate personalized financial information recommendation results for the target user. The determining module is used to perform time alignment, intervention factor identification, and sample normalization on the event impact unit, the market state representation, the decision state representation, and historical exposure interaction records to generate bias estimation input; based on the bias estimation input, it constructs a causal dependency structure between target users, exposure opportunities, event timeliness, and interaction response, and performs counterfactual association estimation, bias weight correction, exposure offset verification, and timeliness decay compensation to generate corresponding causal association estimation results; based on the causal association estimation results, it performs bias-free gain prediction, recommendability assessment, candidate screening, and association ranking on the event impact unit to determine the set of candidate event impact units corresponding to the target user.

Images