User feature and external source double-engine pre-reasoning recommendation method and system

Abstract

The application provides a user feature and external source double-engine pre-reasoning recommendation method and system, relates to the technical field of intelligent recommendation, user intention modeling and information pushing. The method and system fuse user instant context and long-term features to generate an IDR snapshot, simultaneously perform structured analysis on a preset high-value external source to generate a source event object, perform cross-space matching by a double-engine matching module, and write a candidate event into a candidate pool; before pushing, the pre-reasoning recombinant module performs abstract compression, cross-source splicing, conflict resolution and action suggestion generation on the candidate event, and outputs a pre-reasoning recommendation card after cognitive processing; finally, the active pushing arrangement module selects the best pushing strategy in combination with interruption cost and timing window; the upgrade from original material distribution to pre-reasoning cognitive card pushing is realized, the user secondary screening and processing cost is significantly reduced, and information overload is alleviated.

Description

Technical Field

[0001] This invention relates to the fields of intelligent recommendation, user intent modeling, and information push technology, specifically to a dual-engine pre-inference recommendation method and system based on user features and external information sources. Background Technology

[0002] Currently, information retrieval systems mainly employ two technical approaches: one is for users to explicitly input search criteria through a search box or dialog box, and the system then retrieves and returns the results; the other is for the platform to perform collaborative filtering or content tag retrieval based on users' historical behavior data, directly pushing the original articles, posts, or papers to users.

[0003] The above solutions have significant drawbacks: First, search-based systems rely entirely on explicit user triggers; when users only have vague ideas or potential needs, the system cannot proactively perceive and provide relevant information. Second, traditional recommendation systems typically only recommend raw materials, requiring users to read, filter, and process them themselves, resulting in a heavy cognitive burden. Third, traditional recommendation algorithms often rely on historical interests for similarity-based distribution, making it difficult to simultaneously combine users' immediate intentions, long-term characteristics, and dynamic changes from specific high-value external sources, leading to overly broad or delayed recommendation results. Fourth, existing systems generally lack the closed-loop processing capability of "performing cognitive processing and pre-inference after matching before deciding whether to proactively push," failing to transform raw source events into high-value inference result cards for proactive supply. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a dual-engine pre-inference recommendation method and system based on user features and external information sources, which solves the problem of poor information push performance.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A dual-engine pre-inference recommendation method based on user features and external information sources, the method comprising: Continuously collect real-time contextual signals and long-term behavioral characteristics of users to generate user profiles and IDR snapshots; Continuously pull or receive incremental content from a preset set of external information sources and parse it into information source event objects; Calculate candidate matching scores for IDR snapshots and source event objects, and write eligible events into the candidate pool; Perform pre-inference reorganization on a single event or a set of events on the same topic in the candidate pool to form a pre-inference recommendation card and calculate the pre-inference output score; When the pre-inference output score meets the conditions, the corresponding card is actively pushed and arranged. The active push score is calculated by combining the timing, channel and disturbance cost, and it is decided to push immediately, push delayed, push aggregated or keep silent. Send the pre-inference recommendation cards to the user's terminal and record the push history; Collect user feedback on the card and generate feedback events; The matching rules, pre-inference rules, push thresholds, and source trust levels are updated based on feedback events.

[0006] Preferably, the IDR snapshot is used to describe the user's potential information needs, topics of interest, task objectives, risk preferences, timeliness requirements, and content consumption load at the current moment, and the IDR snapshot is dynamically refreshed as the user's real-time signals change.

[0007] Preferably, the information source event object includes: topic tag, timeliness tag, entity tag, evidence citation, credibility level, update type, and actionable field.

[0008] Preferably, the candidate matching score is calculated using the following formula: MatchScore(i,e)=α1×IntentFit(i,e)+α2×TopicOverlap(i,e)+α3×SourceAuthority(e)+α4×Freshness(e)+α5×Actionability(i,e)-α6×RedundancyPenalty(i,e)-α7×NoisePenalty(e); Where i represents the current user's IDR snapshot; e represents a candidate source event; IntentFit(i,e) represents the degree of matching between source event e and the instantaneous intent vector in IDR snapshot i; TopicOverlap(i,e) represents the degree of overlap between source event e and user long-term topic characteristics; SourceAuthority(e) represents the credibility and professionalism of the source to which the information event e belongs; Freshness(e) represents the timeliness of the information source event e; Actionability(i,e) represents the potential for source event e to be transformed into a specific action suggestion; RedundancyPenalty(i,e) represents the duplicate penalty for source event e and recently pushed or consumed content; NoisePenalty(e) represents the penalty for low-quality noise, marketing information, or non-target domain information in source event e; α1~α7 represent the weight coefficients of each matching factor; The candidate entry threshold is set based on historical push hit rate, manually labeled samples, and the cost of allowing disturbances. When MatchScore(i,e) is higher than the candidate entry threshold, the system will write the corresponding source event into the candidate pool. When MatchScore(i,e) is in the observation interval, the system waits for more source events on the same topic or more user context signals before re-evaluating; When MatchScore(i,e) is lower than the threshold, the system only retains the event index and does not enter the active push process.

[0009] Preferably, the pre-reasoning reorganization specifically includes: Perform compressed summaries on a single source event, extracting the subject, facts, change points, affected objects, and actionable fields; Perform cross-source splicing on multiple information source events on the same topic, and merge mutually complementary evidence; Perform trust level comparison and conflict marking for conflicting sources; Based on the current task context in the IDR snapshot, determine whether the source event has a direct impact, indirect impact, or is only of concern to the user's ongoing task; Generate action suggestions, reasons for concern, and risk warnings; Output structured recommendation cards suitable for proactive push notifications.

[0010] Preferably, the pre-inference output score is calculated using the following formula: ReasonScore(c)=β1×RelevanceGain(c)+β2×EvidenceIntegrity(c)+β3×InferenceClarity(c)+β4×ActionPlanValue(c)+β5×TimelinessGain(c)-β6×HallucinationRisk(c)-β7×ConflictPenalty(c); Where c represents the candidate set of cards to be generated; RelevanceGain(c) represents the degree of improvement in user relevance of the card relative to the original material after pre-inference; EvidenceIntegrity(c) represents the completeness of the chain of evidence in the card; InferenceClarity(c) represents the clarity of the inference conclusion; ActionPlanValue(c) represents the executable value of the action suggestion generated by the card for the current task; TimelinessGain(c) represents the timeliness benefit of actively displaying this information in the current time window; HallucinationRisk(c) represents the risk that the model introduces fictitious conclusions during the recombination and reasoning process; ConflictPenalty(c) represents the penalty term when there are still unresolved conflicts within the candidate set; β1~β7 represent the weight coefficients of each factor in the pre-inference; Pre-inference is set based on historical card adoption rates, user subsequent action conversion rates, and the cost of misleading, using thresholds and manual review intervals. When ReasonScore(c) is higher than the pre-inference pass threshold, the system generates a pushable ReasonedCard; When ReasonScore(c) is within the manual review range, the system will only output candidate abstracts or enter delayed review. When ReasonScore(c) is lower than the threshold, the system retains only the original candidates and does not trigger active push.

[0011] Preferably, the active push score is calculated using the following formula: PushScore(r)=γ1×CardValue(r)+γ2×TimingSuitability(r)+γ3×ChannelReachability(r)+γ4×TaskUrgency(r)-γ5×InterruptionPenalty(r)-γ6×FatiguePenalty(r) Where r represents the pre-inference recommendation card to be pushed; CardValue(r) represents the overall value of the card itself; TimingSuitability(r) represents the degree of suitability for pushing the card to the user at the current time. ChannelReachability(r) represents the reachability of the current push channel; TaskUrgency(r) indicates the urgency of the task corresponding to the card; InterruptionPenalty(r) represents the loss caused by interrupting the user in the current state; FatiguePenalty(r) represents the penalty for fatigue caused by frequent push notifications recently; γ1~γ6 represent the weight coefficients of each factor in the push arrangement; The immediate push threshold, delayed push interval, and silent threshold are set based on historical open rate, disturbance tolerance, and task loss function; When PushScore(r) is higher than the immediate push threshold, the system pushes immediately; When PushScore(r) is in the delayed push interval, the system adds the card to the standby queue and waits for the timing window to re-evaluate it; When PushScore(r) is below the silence threshold, the system will only retain the card for users to view later or for subsequent aggregated push notifications.

[0012] A dual-engine pre-inference recommendation system based on user features and external information sources, the system comprising: a multimodal acquisition module, an IDR generation module, an information source access module, an information source parsing module, a dual-engine matching module, a pre-inference reorganization module, an active push orchestration module, a feedback learning module, and a recommendation asset storage module; The multimodal acquisition module is used to acquire the user's real-time signals and long-term characteristics in the terminal environment; The IDR generation module is used to fuse real-time signals and long-term features into a user intent data representation, i.e., an IDR snapshot; The source access module is used to access a preset set of external sources; The source parsing module is used to parse the raw content from external sources into standardized source event objects; The dual-engine matching module is used to perform cross-space matching between IDR snapshots and source event objects, identify candidate information most relevant to the user's current needs, and form candidate matching objects; The pre-inference reorganization module is used to perform summary compression, cross-source evidence splicing, task-related inference, impact scope judgment, action suggestion generation, and risk warning generation on candidate matching objects, and outputs pre-inference recommendation cards; The proactive push orchestration module is used to select a push strategy based on the user's current interruption cost, time window, and card value; The feedback learning module is used to collect user feedback on the pushed cards and update the matching parameters and push threshold. The recommended asset storage module is used to uniformly store IDR snapshots, source event objects, candidate matching objects, pre-inference recommendation cards, push records and feedback events, and supports playback, comparison, versioning and effect evaluation.

[0013] Preferably, the feedback learning module receives feedback of the following types: Click directly to expand and read, save or forward; Ignore, close, silently slide past, or unsubscribe; Based on the card, you can continue to ask a second question, conduct a second search, or make a second edit; Follow the card's suggestions to perform subsequent actions such as code updates, investment research, reading papers, or tracking repositories; The conclusions of the card are evaluated explicitly as useful, useless, misleading, or requiring review.

[0014] Preferably, the feedback learning module updates the user profile, IDR snapshot generation parameters, source object credibility level, candidate matching score weight, pre-inference output score weight, and proactive push score weight based on feedback, and marks high-value card templates, low-quality sources, easily misleading inference paths, and high-fatigue push periods, thereby forming a continuous optimization closed loop of dual-engine proactive recommendation.

[0015] (III) Beneficial Effects This invention provides a dual-engine pre-inference recommendation method and system based on user features and external information sources. Compared with existing technologies, it has the following advantages: In this invention, an IDR snapshot is generated by fusing the user's immediate context with long-term features. Simultaneously, a structured parsing of preset high-value external information sources generates source event objects. A dual-engine matching module performs cross-space matching and writes candidate events into a candidate pool. Before push notifications, a pre-inference reorganization module performs summary compression, cross-source splicing, conflict resolution, and action suggestion generation on the candidate events, outputting cognitively processed pre-inference recommendation cards. Finally, an active push orchestration module selects the optimal push strategy based on interruption costs and timing windows. This upgrade from raw material distribution to pre-inference cognitive card push notifications significantly reduces the user's secondary screening and processing costs, alleviating information overload. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a diagram of the overall system architecture in an embodiment of the present invention.

[0018] Figure 2 This is a flowchart of pre-inference recombination in an embodiment of the present invention.

[0019] Figure 3 This is a flowchart of the method in an embodiment of the present invention.

[0020] Figure 4 This is a state machine diagram for the pre-inference recommendation card in an embodiment of the present invention.

[0021] Figure 5 This is a feedback learning closed-loop diagram in an embodiment of the present invention.

[0022] Figure 6 This is a flowchart of a specific implementation method in Embodiment 1 of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] This application provides a dual-engine pre-inference recommendation method and system based on user features and external information sources, which solves the problem of poor information push performance.

[0025] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods. Example

[0026] like Figures 1-5 As shown, this invention provides a dual-engine pre-inference recommendation method based on user features and external information sources, the method comprising: Continuously collect real-time contextual signals and long-term behavioral characteristics of users to generate user profiles and IDR snapshots; Continuously pull or receive incremental content from a preset set of external information sources and parse it into information source event objects; Calculate candidate matching scores for IDR snapshots and source event objects, and write eligible events into the candidate pool; Perform pre-inference reorganization on a single event or a set of events on the same topic in the candidate pool to form a pre-inference recommendation card and calculate the pre-inference output score; When the pre-inference output score meets the conditions, the corresponding card is actively pushed and arranged. The active push score is calculated by combining the timing, channel and disturbance cost, and it is decided to push immediately, push delayed, push aggregated or keep silent. Send the pre-inference recommendation cards to the user's terminal and record the push history; Collect user feedback on the card and generate feedback events; The matching rules, pre-inference rules, push thresholds, and source trust levels are updated based on feedback events.

[0027] One of the key features of this invention is that it does not wait for the user to input a complete query, but first forms an IDR snapshot, and then performs dual-engine matching between the IDR snapshot and the source event object.

[0028] The IDR snapshot is used to describe the user's potential information needs, topics of interest, task objectives, risk preferences, timeliness requirements, and content consumption load at the current moment, and the IDR snapshot is dynamically refreshed as the user's real-time signals change.

[0029] The information source event object includes: topic tag, timeliness tag, entity tag, evidence citation, credibility level, update type, and actionable field.

[0030] The candidate matching score is calculated using the following formula: MatchScore(i,e)=α1×IntentFit(i,e)+α2×TopicOverlap(i,e)+α3×SourceAuthority(e)+α4×Freshness(e)+α5×Actionability(i,e)-α6×RedundancyPenalty(i,e)-α7×NoisePenalty(e); Where i represents the current user's IDR snapshot, and e represents the candidate source event; IntentFit(i,e) represents the degree of matching between source event e and the immediate intent vector in IDR snapshot i, obtained by the semantic similarity between the current session topic, edited content, recent operation sequence, and event topic tags; TopicOverlap(i,e) represents the degree of overlap between source event e and the user's long-term topic features, obtained by normalizing the intersection of long-term tags, attention list, and entity; SourceAuthority(e) represents the credibility level and professionalism of the source to which source event e belongs, obtained by whitelist level, historical hit effect, and external credibility score; Freshness(e) represents the timeliness of source event e. Freshness is determined by publication time, update time, and topic decay curve; Actionability(i,e) represents the potential for source event e to be transformed into specific action suggestions, determined by event type, change intensity, and the number of fields associated with the current task; RedundancyPenalty(i,e) represents the penalty for duplicate content between source event e and recently pushed or consumed content, determined by similar card comparison and exposure window statistics; NoisePenalty(e) represents the penalty for low-quality noise, marketing information, or non-target domain information in source event e; α1~α7 represent the weight coefficients of each matching factor, configured according to user type, scenario strategy, or historical click conversion statistics, and normalized before use; The candidate entry threshold is set based on the historical push hit rate, manually labeled samples, and the cost of allowing disturbances. When MatchScore(i,e) is higher than the candidate entry threshold, the system writes the corresponding source event into the candidate pool. When MatchScore(i,e) is in the observation range, the system waits for more source events of the same topic or more user context signals before re-evaluating. When MatchScore(i,e) is lower than the threshold, the system only retains the event index and does not enter the active push process.

[0031] Another significant difference between this invention and traditional recommendation systems is that, before candidate information enters the push process, it is not directly displayed as the original post, original paper, or original repository update, but is formed into recommendation cards through a pre-inference reorganization module.

[0032] The pre-inference reorganization specifically includes: Perform compressed summaries on a single source event, extracting the subject, facts, change points, affected objects, and actionable fields; Perform cross-source splicing on multiple information source events on the same topic, and merge mutually complementary evidence; Perform trust level comparison and conflict marking for conflicting sources; Based on the current task context in the IDR snapshot, determine whether the source event has a direct impact, indirect impact, or is only of concern to the user's ongoing task; Generate action suggestions, reasons for concern, and risk warnings; Output structured recommendation cards suitable for proactive push notifications.

[0033] The pre-inference output score is calculated using the following formula: ReasonScore(c)=β1×RelevanceGain(c)+β2×EvidenceIntegrity(c)+β3×InferenceClarity(c)+β4×ActionPlanValue(c)+β5×TimelinessGain(c)-β6×HallucinationRisk(c)-β7×ConflictPenalty(c); Where, c represents the candidate set of cards to be generated; RelevanceGain(c) represents the improvement in user relevance of the card relative to the original material after pre-inference, obtained by the matching increment between the candidate summary and the IDR snapshot; EvidenceIntegrity(c) represents the completeness of the evidence chain in the card, obtained by the number of cited entries, cross-source consistency, and key field coverage; InferenceClarity(c) represents the clarity of the pre-inference conclusion, obtained by the completeness of the conclusion fields, the closure of the logical chain, and the completeness of the explanatory structure; ActionPlanValue(c) represents the executable value of the action suggestion generated by the card for the current task. , is estimated by the number of suggested steps, the strength of binding with the current task, and the expected benefit; TimelinessGain(c) represents the timeliness benefit brought by actively prompting the information in the current time window, which is obtained by matching the event decay curve and the user task window; HallucinationRisk(c) represents the risk of the model introducing fictitious conclusions during the reorganization and inference process, which is obtained by evidence support rate, cross-source verification, and rule constraint checks; ConflictPenalty(c) represents the penalty term when there are still unresolved conflicts within the candidate set; β1~β7 represent the weight coefficients of each factor in the pre-inference, which are configured and normalized according to the industry scenario, inference conservatism, and historical review results; The pre-inference pass threshold and manual review interval are set based on historical card adoption rate, user subsequent action conversion rate, and misleading cost. When ReasonScore(c) is higher than the pre-inference pass threshold, the system generates a pushable ReasonedCard. When ReasonScore(c) is within the manual review interval, the system only outputs candidate summaries or enters delayed review. When ReasonScore(c) is lower than the threshold, the system only retains the original candidates and does not trigger active push.

[0034] The proactive push orchestration does not simply send messages based on candidate scores, but further incorporates the user's current state, device state, and interruption costs to orchestrate the push timing.

[0035] The active push score is calculated using the following formula: PushScore(r)=γ1×CardValue(r)+γ2×TimingSuitability(r)+γ3×ChannelReachability(r)+γ4×TaskUrgency(r)-γ5×InterruptionPenalty(r)-γ6×FatiguePenalty(r) Where r represents the pre-inference recommendation card to be pushed; CardValue(r) represents the comprehensive value of the card itself, obtained by merging MatchScore and ReasonScore; TimingSuitability(r) represents the suitability of pushing the card to the user at the current time, obtained by the user's idle state, schedule context, task stage, and historical opening time period; ChannelReachability(r) represents the reachability of the current push channel, obtained by the terminal online status, notification permissions, client activity, and channel priority; TaskUrgency(r) represents the urgency of the matter corresponding to the card, obtained by the timeliness of the information source, task deadline, and risk level; InterruptionPenalty(r) represents the loss caused by interrupting the user in the current state, obtained by the current editing intensity, meeting status, and historical interruption tolerance; FatiguePenalty(r) represents the penalty for fatigue caused by frequent pushes recently, obtained by the number of pushes in the past time window, the number of consecutive ignores, and the cumulative exposure of similar topics; γ1~γ6 represent the weight coefficients of each factor in push orchestration, configured and normalized according to terminal type, business scenario, or historical delivery conversion data; The immediate push threshold, delayed push interval, and silence threshold are set based on historical open rate, disturbance tolerance, and task loss function. When PushScore(r) is higher than the immediate push threshold, the system pushes the card immediately. When PushScore(r) is within the delayed push interval, the system adds the card to the standby queue and waits for the timing window to re-evaluate it. When PushScore(r) is lower than the silence threshold, the system only retains the card for users to actively view later or for subsequent aggregated pushes.

[0036] This invention provides a dual-engine pre-inference recommendation system based on user features and external information sources. The system includes: a multimodal acquisition module, an IDR generation module, an information source access module, an information source parsing module, a dual-engine matching module, a pre-inference reorganization module, an active push orchestration module, a feedback learning module, and a recommendation asset storage module. The multimodal acquisition module is used to acquire the user's real-time signals and long-term characteristics in the terminal environment; Real-time signals include: current session content, input fragments, browsing context, editing context, time context, task status, and recent operation sequence; Long-term features include: user roles, topic preferences, professional fields, historical clicks, favorites, following list, knowledge graph tags, and long-term goal tags; The IDR generation module is used to fuse real-time signals and long-term features into a user intent data representation, i.e., an IDR snapshot; IDR snapshots are used to describe a user's potential information needs, topics of interest, task objectives, risk preferences, timeliness requirements, and content consumption load at the current moment, and are not limited to the query text explicitly entered by the user. The source access module is used to access a preset set of external sources; External information sources include: specific Twitter bloggers, specific GitHub repositories, paper repositories, technical blogs, announcement streams, subscription news streams, or internal corporate intelligence streams; The information source access module supports timed fetching, Webhook subscription, incremental polling, or event push access for whitelisted information sources; The source parsing module is used to parse original posts, repository updates, paper abstracts, submission records, issue discussions or announcement texts into standardized source event objects, and generate topic tags, timeliness tags, entity tags, evidence citations, credibility levels, update types and actionable fields; The dual-engine matching module is used to perform cross-space matching between IDR snapshots and source event objects, identify candidate information most relevant to the user's current needs, and form candidate matching objects; The pre-inference reorganization module is used to perform summary compression, cross-source evidence splicing, task-related inference, impact scope judgment, action suggestion generation, and risk warning generation on candidate matching objects, and outputs pre-inference recommendation cards; The pre-inference recommendation cards are not simply a list of raw links, but rather a structured cognitive processing result based on the user's current task context; The proactive push orchestration module is used to determine whether to push immediately, delay push, merge push, or simply put the device into the candidate pool for the next round of evaluation, based on the user's current interruption cost, time window, candidate value, repeated exposure, and push channel status. The feedback learning module is used to receive feedback such as user clicks, ignores, favorites, secondary questions, subsequent actions and explicit ratings, and update IDR generation parameters, matching weights, pre-inference trigger thresholds and push throttling strategies. The recommended asset storage module is used to uniformly store IDR snapshots, source event objects, candidate matching objects, pre-inference recommendation cards, push records and feedback events, and supports playback, comparison, versioning and effect evaluation.

[0037] After the system completes the push notification, it continuously collects user feedback on the recommended cards, rather than simply counting click-through rates.

[0038] The feedback learning module receives feedback of the following types: Click directly to expand and read, save or forward; Ignore, close, silently slide past, or unsubscribe; Based on the card, you can continue to ask a second question, conduct a second search, or make a second edit; Follow the card's suggestions to perform subsequent actions such as code updates, investment research, reading papers, or tracking repositories; Explicitly evaluate the card conclusions as useful, useless, misleading, or requiring review; The feedback learning module updates the user profile, IDR snapshot generation parameters, source object credibility level, candidate matching score weight, pre-inference output score weight, and proactive push score weight based on the above feedback. It also marks high-value card templates, low-quality sources, easily misleading inference paths, and high-fatigue push periods, thus forming a continuous optimization closed loop for dual-engine proactive recommendation. Example

[0039] like Figure 6 As shown, this embodiment provides an intelligent research assistant system for developers. The system pre-follows several Twitter tech bloggers, specific GitHub repositories, and a paper database. When a user is reading the code of a model inference framework, editing related design documents, and repeatedly mentions "context caching optimization," "inference acceleration," and "recently implementable solutions" in the dialog window, the multimodal acquisition module generates a corresponding IDRSnapshot. Subsequently, the source access module captures a new version description released by a GitHub repository, a performance analysis of the version by a Twitter blogger, and the latest abstract of a related paper. The dual-engine matching module determines that these SourceEvents are highly relevant to the current IDR and writes them into the candidate pool. The pre-inference reorganization module further extracts conclusions such as "the new version supports incremental cache compression," "suitable for long context inference," and "compatibility limitations exist," splices the evidence chain, and generates a ReasonedCard that "recommends evaluating this version for the current project's test environment." The proactive push orchestration module detects that the user is currently editing a document but not in a meeting, and that the update has high timeliness value, so it proactively pushes the card via a desktop sidebar reminder. After clicking, the user further views the evidence chain and continues to initiate code evaluation tasks based on the card. The system will eventually write this feedback back to improve the quality of subsequent similar push notifications.

[0040] In summary, compared with the prior art, the present invention has the following beneficial effects: 1. In this embodiment of the invention, the IDR generation module integrates and models the user's immediate thoughts, contextual signals and long-term features, thereby getting rid of the passive triggering mode that relies solely on the query, and enabling the system to predict potential information needs before the user has explicitly asked a question.

[0041] 2. In this embodiment of the invention, the system performs structured analysis on preset high-value information sources through an external information source analysis module, enabling the system to continuously obtain targeted and noise-controlled external events from specific bloggers, repositories, and paper databases, rather than indiscriminately scraping raw content from the entire network.

[0042] 3. In this embodiment of the invention, the dual-engine matching mechanism simultaneously considers immediate intent, long-term features, source credibility, timeliness, and repeated exposure, thereby reducing the problems of topic drift, delayed recommendations, and low-value repeated recommendations in traditional recommendation systems.

[0043] 4. In this embodiment of the invention, the pre-reasoning and reorganization module completes summary compression, evidence splicing, task association reasoning, and action suggestion generation before push, so that the recommendation results are upgraded from the distribution of raw materials to structured cards after cognitive processing, reducing the user's secondary understanding and screening costs.

[0044] 5. In this embodiment of the invention, the card value is jointly modeled with interruption cost, timing window and channel accessibility by the active push orchestration module, so as to avoid information disturbance and user fatigue caused by simple high-frequency push.

[0045] 6. In this embodiment of the invention, the feedback learning module continuously updates the source credibility level, matching weight, pre-inference weight, and push threshold, enabling the system to form an evolvable proactive recommendation closed loop for different users and different task scenarios.

[0046] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0047] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 the present invention.

Claims

1. A dual-engine pre-inference recommendation method based on user features and external information sources, characterized in that, The method includes: Continuously collect real-time contextual signals and long-term behavioral characteristics of users to generate user profiles and IDR snapshots; Continuously pull or receive incremental content from a preset set of external information sources and parse it into information source event objects; Calculate candidate matching scores for IDR snapshots and source event objects, and write eligible events into the candidate pool; Perform pre-inference reorganization on a single event or a set of events on the same topic in the candidate pool to form a pre-inference recommendation card and calculate the pre-inference output score; When the pre-inference output score meets the conditions, the corresponding card is actively pushed and arranged. The active push score is calculated by combining the timing, channel and disturbance cost, and it is decided to push immediately, push delayed, push aggregated or keep silent. Send the pre-inference recommendation cards to the user's terminal and record the push history; Collect user feedback on the card and generate feedback events; The matching rules, pre-inference rules, push thresholds, and source trust levels are updated based on feedback events.

2. The dual-engine pre-inference recommendation method based on user features and external information sources as described in claim 1, characterized in that, The IDR snapshot is used to describe the user's potential information needs, topics of interest, task objectives, risk preferences, timeliness requirements, and content consumption load at the current moment, and the IDR snapshot is dynamically refreshed as the user's real-time signals change.

3. The dual-engine pre-inference recommendation method based on user features and external information sources as described in claim 1, characterized in that, The information source event object includes: topic tag, timeliness tag, entity tag, evidence citation, credibility level, update type, and actionable field.

4. The dual-engine pre-inference recommendation method based on user features and external information sources as described in claim 1, characterized in that, The candidate matching score is calculated using the following formula: MatchScore(i,e)=α1×IntentFit(i,e)+α2×TopicOverlap(i,e)+α3×SourceAuthority(e)+α4×Freshness(e)+α5×Actionability(i,e)-α6×RedundancyPenalty(i,e)-α7×NoisePenalty(e); Where i represents the current user's IDR snapshot; e represents a candidate source event; IntentFit(i,e) represents the degree of matching between source event e and the instantaneous intent vector in IDR snapshot i; TopicOverlap(i,e) represents the degree of overlap between source event e and user long-term topic characteristics; SourceAuthority(e) represents the credibility and professionalism of the source to which the information event e belongs; Freshness(e) represents the timeliness of the information source event e; Actionability(i,e) represents the potential for source event e to be transformed into a specific action suggestion; RedundancyPenalty(i,e) represents the duplicate penalty for source event e and recently pushed or consumed content; NoisePenalty(e) represents the penalty for low-quality noise, marketing information, or non-target domain information in source event e; α1~α7 represent the weight coefficients of each matching factor; The candidate entry threshold is set based on historical push hit rate, manually labeled samples, and the cost of allowing disturbances. When MatchScore(i,e) is higher than the candidate entry threshold, the system will write the corresponding source event into the candidate pool. When MatchScore(i,e) is in the observation interval, the system waits for more source events on the same topic or more user context signals before re-evaluating; When MatchScore(i,e) is lower than the threshold, the system only retains the event index and does not enter the active push process.

5. The dual-engine pre-inference recommendation method based on user features and external information sources as described in claim 1, characterized in that, The pre-inference reorganization specifically includes: Perform compressed summaries on a single source event, extracting the subject, facts, change points, affected objects, and actionable fields; Perform cross-source splicing on multiple information source events on the same topic, and merge mutually complementary evidence; Perform trust level comparison and conflict marking for conflicting sources; Based on the current task context in the IDR snapshot, determine whether the source event has a direct impact, indirect impact, or is only of concern to the user's ongoing task; Generate action suggestions, reasons for concern, and risk warnings; Output structured recommendation cards suitable for proactive push notifications.

6. The dual-engine pre-inference recommendation method based on user features and external information sources as described in claim 1, characterized in that, The pre-inference output score is calculated using the following formula: ReasonScore(c)=β1×RelevanceGain(c)+β2×EvidenceIntegrity(c)+β3×InferenceClarity(c)+β4×ActionPlanValue(c)+β5×TimelinessGain(c)-β6×HallucinationRisk(c)-β7×ConflictPenalty(c); Where c represents the candidate set of cards to be generated; RelevanceGain(c) represents the degree of improvement in user relevance of the card relative to the original material after pre-inference; EvidenceIntegrity(c) represents the completeness of the chain of evidence in the card; InferenceClarity(c) represents the clarity of the inference conclusion; ActionPlanValue(c) represents the executable value of the action suggestion generated by the card for the current task; TimelinessGain(c) represents the timeliness benefit of actively displaying this information in the current time window; HallucinationRisk(c) represents the risk that the model introduces fictitious conclusions during the recombination and reasoning process; ConflictPenalty(c) represents the penalty term when there are still unresolved conflicts within the candidate set; β1~β7 represent the weight coefficients of each factor in the pre-inference; Pre-inference is set based on historical card adoption rates, user subsequent action conversion rates, and the cost of misleading, using thresholds and manual review intervals. When ReasonScore(c) is higher than the pre-inference pass threshold, the system generates a pushable ReasonedCard; When ReasonScore(c) is within the manual review range, the system will only output candidate abstracts or enter delayed review. When ReasonScore(c) is lower than the threshold, the system retains only the original candidates and does not trigger active push.

7. The dual-engine pre-inference recommendation method based on user features and external information sources as described in claim 1, characterized in that, The active push score is calculated using the following formula: PushScore(r)=γ1×CardValue(r)+γ2×TimingSuitability(r)+γ3×ChannelReachability(r)+γ4×TaskUrgency(r)-γ5×InterruptionPenalty(r)-γ6×FatiguePenalty(r) Where r represents the pre-inference recommendation card to be pushed; CardValue(r) represents the overall value of the card itself; TimingSuitability(r) represents the degree of suitability for pushing the card to the user at the current time. ChannelReachability(r) represents the reachability of the current push channel; TaskUrgency(r) indicates the urgency of the task corresponding to the card; InterruptionPenalty(r) represents the loss caused by interrupting the user in the current state; FatiguePenalty(r) represents the penalty for fatigue caused by frequent push notifications recently; γ1~γ6 represent the weight coefficients of each factor in the push arrangement; The immediate push threshold, delayed push interval, and silent threshold are set based on historical open rate, disturbance tolerance, and task loss function; When PushScore(r) is higher than the immediate push threshold, the system pushes immediately; When PushScore(r) is in the delayed push interval, the system adds the card to the standby queue and waits for the timing window to re-evaluate it; When PushScore(r) is below the silence threshold, the system will only retain the card for users to view later or for subsequent aggregated push notifications.

8. A dual-engine pre-inference recommendation system based on user features and external information sources, characterized in that, The system includes: a multimodal acquisition module, an IDR generation module, a source access module, a source parsing module, a dual-engine matching module, a pre-inference reorganization module, an active push orchestration module, a feedback learning module, and a recommendation asset storage module; The multimodal acquisition module is used to acquire the user's real-time signals and long-term characteristics in the terminal environment; The IDR generation module is used to fuse real-time signals and long-term features into a user intent data representation, i.e., an IDR snapshot; The source access module is used to access a preset set of external sources; The source parsing module is used to parse the raw content from external sources into standardized source event objects; The dual-engine matching module is used to perform cross-space matching between IDR snapshots and source event objects, identify candidate information most relevant to the user's current needs, and form candidate matching objects; The pre-inference reorganization module is used to perform summary compression, cross-source evidence splicing, task-related inference, impact scope judgment, action suggestion generation, and risk warning generation on candidate matching objects, and outputs pre-inference recommendation cards; The proactive push orchestration module is used to select a push strategy based on the user's current interruption cost, time window, and card value; The feedback learning module is used to collect user feedback on the pushed cards and update the matching parameters and push threshold. The recommended asset storage module is used to uniformly store IDR snapshots, source event objects, candidate matching objects, pre-inference recommendation cards, push records and feedback events, and supports playback, comparison, versioning and effect evaluation.

9. The dual-engine pre-inference recommendation system based on user features and external information sources as described in claim 8, characterized in that, The feedback learning module receives feedback of the following types: Click directly to expand and read, save or forward; Ignore, close, silently slide past, or unsubscribe; Based on the card, you can continue to ask a second question, conduct a second search, or make a second edit; Follow the card's suggestions to perform subsequent actions such as code updates, investment research, reading papers, or tracking repositories; The conclusions of the card are evaluated explicitly as useful, useless, misleading, or requiring review.

10. The dual-engine pre-inference recommendation method based on user features and external information sources as described in claim 9, characterized in that, The feedback learning module updates user profiles, IDR snapshot generation parameters, source object credibility level, candidate matching score weight, pre-inference output score weight, and proactive push score weight based on feedback. It also marks high-value card templates, low-quality sources, easily misleading inference paths, and high-fatigue push periods, thus forming a continuous optimization closed loop for dual-engine proactive recommendation.

Images