A member preference modeling and commodity recommendation method based on full-channel consumption track

Abstract

This invention discloses a method for member preference modeling and product recommendation based on omnichannel consumption trajectories, relating to the fields of retail digitization and intelligent recommendation. The method includes: collecting member channel consumption records and generating a unified consumption event sequence; generating a unified member index based on identity clues; performing continuity judgment on the unified consumption event sequence based on the unified member index, outputting cross-channel intent continuity coefficients and consumption trajectory fragments; performing preference deposition on the consumption trajectory fragments to generate dynamic member preferences; collecting product operation records and generating product constraint information; and generating product recommendation results based on consumption trajectory fragments, product constraint information, and dynamic member preferences. By unifying omnichannel consumption events, constructing continuous consumption trajectories, and integrating product operation constraints, the method achieves integrated processing of member preference modeling and product recommendation.

Description

Technical Field

[0001] This invention relates to the field of retail digitalization and intelligent recommendation, specifically a method for member preference modeling and product recommendation based on omnichannel consumption patterns. Background Technology

[0002] With the continued development of chain retail, instant retail, and platform retail, member consumption behavior has gradually expanded from single-store scenarios to omnichannel scenarios involving POS terminals, online transaction portals, platform interfaces, and fulfillment chains. Members' browsing, ordering, payment, verification, and fulfillment behaviors across different channels accumulate continuously. Retail enterprises typically manage relevant data through membership systems, order systems, inventory systems, and promotional systems, and attempt to conduct member profiling analysis and product recommendations based on historical consumption records.

[0003] In existing technologies, consumption records generated from different channels generally suffer from inconsistencies in terms of time base, product coding, event type, and identity identification. The behavior of the same member across different channels is difficult to accurately categorize, resulting in unclear sequential relationships between consumption events and hindering the formation of a continuous and reliable consumption trajectory. Simultaneously, product inventory, pricing, promotions, and fulfillment status are scattered across different operational records, and there is a lack of a unified method for linking product constraint information with member consumption trajectories. Consequently, the product recommendation process struggles to simultaneously consider both genuine member preferences and product operational constraints.

[0004] In this context, existing recommendation methods typically rely on static historical orders or single interaction records for preference judgments. These methods struggle to reflect the continuous intent across cross-channel behaviors and handle business relationships such as specification changes, packaging correspondence, combination associations, and fulfillment coordination. This results in insufficient accuracy in member preference modeling and inadequate basis for product recommendations. Therefore, a method for member preference modeling and product recommendation based on omnichannel consumption trajectories is needed to uniformly process multi-source consumption events, identify the continuity of member behaviors, and generate product recommendation results in conjunction with product operation constraints. Summary of the Invention

[0005] Based on the shortcomings of the existing technology described above, the purpose of this invention is to provide a method for member preference modeling and product recommendation based on omnichannel consumption patterns to solve the aforementioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for member preference modeling and product recommendation based on omnichannel consumption patterns, comprising:

[0007] Collect members' channel consumption records, perform time correction, product mapping and event normalization on the channel consumption records, and generate a unified consumption event sequence carrying identity clues and event context information;

[0008] A unified member index is generated by performing association parsing based on identity clues in the unified consumption event sequence;

[0009] Based on the unified member index, the unified consumption event sequence is rearranged, and the event context information is judged to determine the continuity. The cross-channel intent continuity coefficient and consumption trajectory fragment are output.

[0010] Execute preference deposition on consumption trajectory segments, and generate dynamic member preferences by combining cross-channel intent acceptance coefficients;

[0011] Collect product operation records, and organize and associate the execution status of product operation records to generate product constraint information;

[0012] Based on consumption trajectory fragments and product constraint information, correlation filtering is performed, and product recommendation results are generated by combining dynamic member preferences.

[0013] The present invention is further configured such that generating a unified consumption event sequence carrying identity clues and event context information includes:

[0014] Channel consumption records include cashier records, online transaction records, platform interface records, and fulfillment records;

[0015] The channel consumption records are decomposed into fields, and the original time information, channel source information, store information, product information, action information, order association information, payment association information, promotion information, fulfillment association information and identity information are extracted and encapsulated according to the preset field structure to generate the original event carrier;

[0016] Based on order association information, payment association information, and performance association information, a unified timestamp is generated by aligning and offsetting the execution time of the original event carrier.

[0017] Perform matching constraints and hierarchical validations on product information against a pre-defined standard product database to generate a unified product index;

[0018] By combining action information, order association information, payment association information, and fulfillment association information, the execution events of the original event carrier are normalized to generate a unified event type. The channel source information, store information, promotion information, fulfillment association information, and unified timestamp are organized to generate event context information.

[0019] Based on the identity information, a reliability screening process is performed to generate identity clues;

[0020] A unified consumption event sequence is generated by encapsulating and sorting unified timestamps, unified product indexes, unified event types, event context information, and identity clues.

[0021] The present invention is further configured such that the step of generating a unified member index by performing association parsing based on identity clues in a unified consumption event sequence includes:

[0022] Based on the identity clues in the unified consumption event sequence, type sorting, format standardization, and semantic regularization are performed to generate identity clue groups divided by telephone number identifier, payment identifier, device identifier, account identifier, and address identifier;

[0023] By combining the event context information in the unified consumption event sequence, source reliability verification, format integrity verification, and context consistency verification are performed on the identity clue group to generate the event-level identity description of the corresponding unified consumption event;

[0024] Based on identity clue groups and event-level identity descriptions, candidate association filtering and association scoring are performed on consumption events in a unified consumption event sequence to generate a candidate association set;

[0025] Merge and aggregate the candidate association set and perform conflict checks to generate member merge groups. Then, perform index allocation on the member merge groups to generate a unified member index.

[0026] The present invention is further configured such that, in rearranging the unified consumption event sequence based on the unified member index and performing succession discrimination on the event context information, the following steps are included:

[0027] Based on the member merging groups corresponding to the unified member index, the unified consumption event sequence is divided by affiliation and sorted by time sequence to generate member event subsequences;

[0028] Perform candidate construction and span filtering on adjacent consumption events in the member event subsequence to generate valid successor candidate event pairs;

[0029] Extract the event context information of the preceding and subsequent consumption events in the valid candidate event pair, perform time interval calculation, channel migration identification, promotion link comparison and fulfillment status connection determination, and generate a context description.

[0030] Extract the unified product index, unified event type, and store information of the preceding and subsequent consumption events in the valid candidate event pair, perform product correspondence comparison, event flow order determination, and store acceptance determination, and generate a continuous semantic description of the event.

[0031] The present invention is further configured such that the output of cross-channel intent reception coefficient and consumption trajectory fragment includes:

[0032] A joint determination is performed on the context description and the continuous semantic description of the event to generate the cross-channel intent acceptance coefficient for the effective acceptance candidate event pair;

[0033] Based on the cross-channel intent acceptance coefficient, the execution order of adjacent consumption events in the member event subsequence is segmented to generate initial consumption trajectory fragments;

[0034] Perform internal continuity checks and boundary adjustments on the initial consumption trajectory segment to generate a new consumption trajectory segment.

[0035] The present invention is further configured such that the step of performing preference deposition on consumption trajectory segments and generating dynamic member preferences by combining cross-channel intent reception coefficients includes:

[0036] Based on consumption events in consumption trajectory segments, a unified product index, a unified event type, and event context information are extracted, and preference semantic mapping is performed to generate event preference primitives.

[0037] By combining cross-channel intent acceptance coefficients, unified event types, and time information, the event deposition intensity is used to determine the generated deposition weights.

[0038] The event preference primitives and deposition weights are sequentially deposited to generate fragment preference representations, and the cross-channel intent acceptance coefficients within the consumption trajectory fragments are organized to generate fragment acceptance stability information.

[0039] Based on fragment preference representation, fragment-inherited stable information and temporal information, cross-fragment fusion is performed to generate long-term and short-term preferences;

[0040] Dynamic preferences for members are generated by dynamically fusing long-term and short-term preferences.

[0041] The present invention is further configured such that the step of collecting product operation records, normalizing and associating the execution status of product operation records to generate product constraint information includes:

[0042] Collect product operation records and extract product identifiers, specification information, channel operation identifiers, store operation identifiers, operation time, inventory information, price information, promotion execution information, fulfillment execution information, and source information to generate product operation atoms;

[0043] Product identification and specification information are mapped to a unified product index space. Time window slicing is performed on the operation time. Based on the unified product index, channel operation identification, store operation identification and slice time, the atomic execution of product operation is aggregated to generate scenario product units.

[0044] The execution status of inventory information, price information, promotion execution information and fulfillment execution information in the scene product unit is unified, and the execution status quantity is generated by merging and integrating the source information.

[0045] Based on the specification relationships, packaging relationships, combination relationships and fulfillment connection relationships corresponding to the unified product index, the scene product units are associated and organized to generate associated description quantities.

[0046] The unified product index, channel operation identifier, store operation identifier, slice time, scene state quantity, and associated description quantity are encapsulated to generate product constraint information.

[0047] The present invention is further configured such that the association filtering based on consumption trajectory fragments and product constraint information includes:

[0048] Extract the unified product index, unified event type, and event context information from the consumption trajectory segment, and match them with the unified product index, channel operation identifier, store operation identifier, and slice time in the product constraint information to generate candidate product items;

[0049] Based on the inventory status, price status, promotion status, and fulfillment status in the scenario state quantity, constraint filtering is performed on candidate product items to generate an initial candidate product set;

[0050] Based on the specification relationships, packaging relationships, combination relationships, and fulfillment connection relationships in the associated descriptive quantities, the initial candidate product set is expanded and deduplicated to generate a new candidate product set.

[0051] The present invention is further configured such that the step of generating product recommendation results by combining member dynamic preferences includes:

[0052] Perform preference matching between members' dynamic preferences and the candidate product set to generate preference matching results corresponding to the candidate products;

[0053] By combining consumption trajectory fragments and cross-channel intent reception coefficients, the execution trajectory support for the candidate product set is determined, and the trajectory support results corresponding to the candidate products are generated.

[0054] Based on preference matching results and trajectory support results, the candidate product set is recommended and ranked to generate product recommendation results;

[0055] Based on the feedback records corresponding to the product recommendation results, update the cross-channel intent reception coefficient and member dynamic preferences.

[0056] The present invention is further configured such that updating the cross-channel intent reception coefficient and member dynamic preferences based on the feedback records corresponding to the product recommendation results includes:

[0057] Collect feedback records corresponding to product recommendation results, extract exposure information, click information, add-to-cart information, order information, redemption information, refund information, and ignore information, and generate a feedback event sequence;

[0058] The feedback event sequence is correlated with product recommendation results, consumption trajectory fragments, and cross-channel intent reception coefficients to generate feedback correlation results.

[0059] Based on the feedback correlation results, the cross-channel intent acceptance coefficient is corrected to generate an updated cross-channel intent acceptance coefficient;

[0060] Based on the updated cross-channel intent reception coefficient and feedback correlation results, the member dynamic preferences are updated to generate updated member dynamic preferences.

[0061] This invention provides a method for member preference modeling and product recommendation based on omnichannel consumption trajectories. The method collects members' channel consumption records, performs time correction, product mapping, and event normalization on these records, generating a unified consumption event sequence carrying identity clues and event context information. It then performs association parsing based on the identity clues in the unified consumption event sequence to generate a unified member index. Based on the unified member index, it rearranges the unified consumption event sequence, performs continuity judgment on the event context information, and outputs cross-channel intent continuity coefficients and consumption trajectory fragments. It performs preference deposition on the consumption trajectory fragments and generates dynamic member preferences by combining the cross-channel intent continuity coefficients. It collects product operation records, performs state normalization and association organization on these records to generate product constraint information, and performs association filtering based on the consumption trajectory fragments and product constraint information, combining the dynamic member preferences to generate product recommendation results. The beneficial effects include:

[0062] 1. Improved integrity of consumption trajectory construction: By correcting the execution time of channel consumption records, mapping products, unifying events and parsing identity associations, consumption events scattered across different channels are unified under the same member index. Combined with the acceptance judgment, consumption trajectory fragments are formed, giving cross-channel consumption behavior a continuous organizational foundation.

[0063] 2. Enhanced Refinement of Member Preference Modeling: By introducing consumption trajectory fragments and cross-channel intent reception coefficients into the preference deposition process, the contribution of different consumption events to preference formation is differentiated, and dynamic fusion of long-term and short-term preferences is implemented, making the description of member preferences more consistent with the actual consumption evolution process.

[0064] 3. Improved synergy between recommendation results and product constraints: By normalizing and associating the execution status of product operation records, product constraint information is formed, including inventory, price, promotion, fulfillment, and product relationships. This information is then combined with consumption trajectory fragments and dynamic member preferences for joint screening, thereby improving the consistency of constraints in the product recommendation process.

[0065] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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. In the drawings:

[0067] Figure 1 The flowchart illustrates a method for member preference modeling and product recommendation based on omnichannel consumption patterns, as an exemplary embodiment of the present invention. Detailed Implementation

[0068] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0069] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0070] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0071] Example 1:

[0072] A method for member preference modeling and product recommendation based on omnichannel consumption patterns, such as... Figure 1 As shown, it includes:

[0073] Collect members' channel consumption records, perform time correction, product mapping and event normalization on the channel consumption records, and generate a unified consumption event sequence carrying identity clues and event context information;

[0074] A unified member index is generated by performing association parsing based on identity clues in the unified consumption event sequence;

[0075] Based on the unified member index, the unified consumption event sequence is rearranged, and the event context information is judged to determine the continuity. The cross-channel intent continuity coefficient and consumption trajectory fragment are output.

[0076] Execute preference deposition on consumption trajectory segments, and generate dynamic member preferences by combining cross-channel intent acceptance coefficients;

[0077] Collect product operation records, and organize and associate the execution status of product operation records to generate product constraint information;

[0078] Based on consumption trajectory fragments and product constraint information, correlation filtering is performed, and product recommendation results are generated by combining dynamic member preferences.

[0079] The present invention is further configured such that generating a unified consumption event sequence carrying identity clues and event context information includes:

[0080] Channel consumption records include POS records, online transaction records, platform interface records, and fulfillment records. Specifically, first, access business records from different sources that are all directly related to the consumption process. Channel consumption records include POS records, online transaction records, platform interface records, and fulfillment records. Among them, POS records refer to transaction records output by offline POS terminals, which at least include the transaction time, product code, transaction price, store identification, and payment status; online transaction records refer to order records generated by online transaction portals, which at least include the order time, product identification, order status, promotional information, and member account identification; platform interface records refer to transaction data transmitted by external platforms through interfaces, which at least include the platform order number, product description, payment transaction record, and delivery method; fulfillment records refer to transaction data related to the order... The fulfillment status data corresponding to a single completion process includes at least the verification status, outbound status, delivery status, receipt status, or self-pickup completion status. During data collection, a collection period, field whitelist, and deduplication rules are first set for each type of record. The collection period is used to constrain the time range for record retrieval or reception. The field whitelist is used to limit the fields to be retained, preventing irrelevant fields from entering subsequent processing. The deduplication rules are used to remove duplicate records based on order identifier, payment identifier, and timestamp. After this step, a set of original records with clear sources and defined field ranges is formed, providing input for subsequent unified processing. The effect of this step is to incorporate data originally scattered from different sources into a single processing entry point, and to avoid the same consumption process being counted repeatedly through source classification and deduplication control.

[0081] The channel consumption records are decomposed into fields to extract original time information, channel source information, store information, product information, action information, order association information, payment association information, promotion information, fulfillment association information, and identity information. These are then encapsulated according to a preset field structure to generate original event carriers. Specifically, each collected record is broken down into basic fields and encapsulated into an original event carrier according to a preset field structure. The original time information comes from the valid time fields of the transaction occurrence time, payment completion time, and fulfillment status update time. Channel source information identifies whether the record comes from the cashier, online transaction, platform interface, or fulfillment end. Store information comes from the store number, fulfillment store number, or warehouse / distribution node identifier. Product information comes from the product code, product name, specifications, and packaging description. Action information comes from the original business action name, such as placing an order, payment, verification, refund, or fulfillment completion. Order association information comes from the order number, sub-order number, refund number, and verification number. Payment-related information comes from payment transaction number, payment channel identifier, and payment status; promotional information comes from discount activity identifier, coupon redemption identifier, membership benefit usage identifier, or discount rule identifier; fulfillment-related information comes from outbound status, delivery status, self-pickup status, and receipt status; identity information comes from member number, mobile phone number, payment account fragment, device identifier, or address fragment; the preset field structure is not a simple list, but predefines field positions, field types, required fields, and fault-tolerant fields to ensure that records from different sources can be read according to the same structure; for missing fields, they are not directly discarded, but a null value marker is retained, and the missing type is recorded for easy reliability determination later; after encapsulation, each record is transformed into a raw event carrier; the effect of this step is to transform heterogeneous records into homogeneous data objects, making data from different sources comparable; at the same time, through null value markers and field type markers, sufficient basis is retained for subsequent time correction, event normalization, and identity filtering;

[0082] Based on order association information, payment association information, and fulfillment association information, a unified timestamp is generated by performing time alignment and offset correction on the original event carrier. Specifically, since the time base, upload latency, and write time points of different source systems are not consistent, directly using the original time information will lead to misalignment in the time sequence of the same transaction process. In this embodiment, the order association information, payment association information, and fulfillment association information are first used to find event pairs belonging to the same transaction link. For example, if a payment completion record and an order confirmation record correspond to the same order number, they are considered as one set of aligned samples; if a verification record and a fulfillment completion record correspond to the same verification order number, they are considered as another set of aligned samples. During alignment, the time offset level of a large number of samples within a certain period of time under the same source is accumulated and statistically analyzed, and the median difference is preferentially taken as the basic offset. The median difference is used to avoid abnormally delayed records having an excessive impact on the overall offset result. After the basic offset is determined, local correction is performed based on the time proximity between the original event carrier and the adjacent aligned samples. Local correction means prioritizing the alignment sample that is closer in time for a certain record to eliminate local offsets caused by batch backhaul, network jitter, and interface buffering. After completing the basic offset and local correction, a unified timestamp is generated. The source of the unified timestamp is the result of the original time information after time alignment and offset correction, and its value always falls under a unified time base. The effect of this step is to restore the true chronological relationship of events within the same consumption process, avoid the time sequence reversal phenomenon such as payment completion before order placement or fulfillment completion before payment confirmation, and provide a reliable time basis for subsequent sorting and acceptance judgment.

[0083] The product information is matched against a pre-defined standard product database, and hierarchical verification is performed to generate a unified product index. Specifically, in this embodiment, product information is not directly equivalent to a single field, but is composed of product code, product name, specification description, and packaging description. The pre-defined standard product database is a pre-established standardized product master data set, which includes at least a standard product index, standard product name, specification hierarchy, packaging hierarchy, and product relationship table. Specification hierarchy refers to the hierarchical relationship of the same product under different capacities, weights, or quantities of packaging. Packaging hierarchy refers to the affiliation relationship between single items, combined items, and outer packaging. During matching, the product code is used as the primary matching condition. When the product code is missing or the coding systems from different sources are inconsistent, the product name, specification description, and packaging description are introduced for auxiliary comparison. The comparison adopts a constraint matching method, which requires that the product names are similar, the specification descriptions are compatible, and the packaging descriptions do not conflict. Only when all three conditions are met can the product enter the candidate range. After the candidate range is formed, hierarchical verification is performed. The purpose of hierarchical verification is to prevent combined products from being mistakenly mapped to single items and to prevent large packages from being mistakenly mapped to small specifications. During verification, the system checks whether the specification description in the current record is consistent with the specification level of the candidate standard product, and whether the packaging description is consistent with the packaging level of the candidate standard product. It also confirms whether there is a convertible relationship or inclusion relationship by referring to the product relationship table. After matching constraints and level verification, a unique unified product index is determined. If the verification fails, the product is retained as a product to be confirmed and is not directly included in the unified product index results. The effect of this step is to unify products from different sources, different coding systems, and different description methods into the standard product space, so as to avoid the subsequent fragmentation of the same product for statistics or the mixing of different products.

[0084] By combining action information, order association information, payment association information, and fulfillment association information, the original event carrier is normalized to generate a unified event type. Channel source information, store information, promotional information, fulfillment association information, and a unified timestamp are organized to generate event context information. Specifically, event normalization refers to mapping business actions from different sources and with different naming methods to a standard event type. The unified event type includes at least order placement, payment, verification, refund, fulfillment completion, and transaction cancellation. When performing event normalization, not only is the action information itself examined, but also order association information, payment association information, and fulfillment association information are considered for joint judgment. For example, if the action name is "Transaction Completed," but the payment status is not completed and the fulfillment status is only "Outbound," it cannot be directly classified as fulfillment completion. If the action name is "Verified," and the verification order number is valid and the fulfillment status is "Self-Pickup Completed," it is classified as verification. Joint judgment refers to using the action name as the initial candidate and considering the order status, order status, and other relevant information as factors. Payment status and fulfillment status are used as verification conditions to eliminate incompatible event types one by one, ultimately retaining a unique and unified event type. After event unification, event context information is then generated. This event context information is not simply a concatenation of fields, but rather an encapsulation of background elements required for subsequent judgments, including at least channel source information, store information, promotional information, fulfillment association information, and a unified timestamp. Promotional information identifies whether the current event participates in a promotional activity, uses a coupon, or utilizes membership benefits. Fulfillment association information identifies the current fulfillment stage of the event, such as pending shipment, in delivery, self-pickup completed, or receipt completed. After this process, event context information is generated. The benefits of this step are twofold: firstly, it creates a unified event type, allowing actions from different sources to be compared at the same semantic level; secondly, it generates event context information, providing sufficient background evidence for subsequent judgments about whether adjacent events belong to the same consumer intent.

[0085] Identity clues are generated based on the reliability screening performed on the identity information. Specifically, the sources of identity information are wide-ranging, and may include membership numbers, mobile phone numbers, payment account fragments, device identifiers, and address fragments. The distinguishability and stability of different identity information are not consistent. Therefore, in this embodiment, the identity information is first categorized and standardized in format. Category categorization refers to grouping fields with different names but the same essence into the same category, such as grouping membership number and membership card number into the membership account category. Format standardization refers to removing spaces, unifying capitalization, unifying delimiters, and retaining rules. Then, reliability screening is performed on each piece of identity information. Reliability screening considers at least three types of factors: the first type is source reliability, that is, in what source record does the identity information appear, usually in the payment account of the payment completion record. The first category is user fragments, which are more stable than manually entered notes; the second category is completeness, i.e., whether the information is missing key fragments or is just overly anonymized content; the third category is contextual consistency, i.e., whether the identity information is consistent with the current store, channel, time, and fulfillment chain; identity information that simultaneously meets the requirements of reliable source, complete format, and consistent context is retained as identity clues; information that only meets some of the conditions is retained as low-confidence markers; information with obvious conflicts is directly eliminated; identity clues refer to the set of identity markers that can be used for subsequent member association parsing after screening, rather than the original identity fields themselves; the effect of this step is to converge the noisy original identity information into reliable clues that can be used for member merging, and to avoid erroneous identity information interfering with the generation of a unified member index;

[0086] A unified consumption event sequence is generated by encapsulating and sorting unified timestamps, unified product indexes, unified event types, event context information, and identity clues. Specifically, in this embodiment, the unified timestamps, unified product indexes, unified event types, event context information, and identity clues obtained in the aforementioned steps are integrated into a unified consumption event. The unified consumption event is the basic data unit for all subsequent processing, and it retains at least five types of content: time attributes, product attributes, event attributes, context attributes, and identity attributes. After encapsulation, sorting is also required. The sorting is first performed in ascending order according to the unified timestamp to restore the unified time order of all consumption events. When unified timestamps are the same or very close, the order is further determined according to the preset business priority. The process involves sequential execution and concurrency resolution. Concurrency resolution refers to determining the order of several events occurring at the same point in time according to business logic. For example, order confirmation is usually done first, followed by payment confirmation, and then fulfillment confirmation. The preset business priority order is derived from standard transaction chain rules, rather than being arbitrarily specified. After sorting, a unified consumption event sequence is generated. The meaning of a unified consumption event sequence is an ordered set of consumption events organized according to a unified time base, a unified product space, and a unified event semantics. Each event carries identity clues and event context information. The effect of this step is to provide a unified, continuous, and traceable data foundation for subsequent identity association parsing, acceptance judgment, and preference deposition, and to eliminate the uncertainty caused by differences in sources to subsequent analysis through encapsulation and sorting.

[0087] The present invention is further configured such that the step of generating a unified member index by performing association parsing based on identity clues in a unified consumption event sequence includes:

[0088] Based on the identity clues in the unified consumption event sequence, type categorization, format standardization, and semantic normalization are performed to generate identity clue groups divided into telephone number identifiers, payment identifiers, device identifiers, account identifiers, and address identifiers. Specifically, the identity clues carried by each consumption event in the unified consumption event sequence are first read, and the identity clues are initially split according to the source field. The telephone number identifier comes from the mobile phone number, contact number, recipient's phone number, or reserved phone number field; the payment identifier comes from the payment account fragment, payment transaction association number, or payment channel account mark; the device identifier comes from the terminal number, device fingerprint, login device number, or communication device mark; the account identifier comes from the member number, registered account, platform user number, or card number; and the address identifier comes from the delivery address fragment, billing address fragment, or fulfillment address fragment. After the source identification is completed, type categorization is performed on various identity clues. Type categorization refers to grouping fields with different names but the same meaning into the same category to eliminate differences in field naming between different source systems; then, format standardization is performed. Format standardization includes removing spaces and separators, unifying the representation of numbers and characters, unifying capitalization, unifying mask formats, and unifying truncation lengths. For example, phone numbers retain only the standard digits and remove spaces and connectors; address segments are split into levels such as province, city, district, street, and house number, and stable segments are retained; payment identifiers retain stable prefixes, stable middle segments, or stable end segments according to platform preset rules. Then, semantic regularization is performed. Semantic regularization refers to unifying the semantics of clues with similar meanings but different expressions within the same type. For example, aligning all phone numbers and anonymized phone numbers to the same comparison template, and converting device identifiers containing prefix codes into a unified device expression. After the above processing, identity clue groups are obtained, divided by phone number identifier, payment identifier, device identifier, account identifier, and address identifier. The effect of this step is to eliminate the differences in field names, format styles, and semantic expressions of data from different sources, so that subsequent reliability verification and event association are based on unified clues, avoiding the splitting of the same member due to inconsistent formats, or the erroneous merging of different members due to semantic confusion.

[0089] Combining the event context information in the unified consumption event sequence, source reliability verification, format integrity verification, and context consistency verification are performed on the identity clue groups to generate event-level identity descriptions corresponding to the unified consumption events. Specifically, the event context information corresponding to each consumption event is read. The event context information comes from the unified consumption events formed in the preceding steps, including channel source information, store information, promotion information, fulfillment association information, and unified timestamps. Subsequently, source reliability verification is performed on the identity clue groups. Source reliability verification refers to determining the credibility of an identity clue based on the record link and field position in which it appears. Identity clues directly from payment completion, account login, reconciliation confirmation, or completed fulfillment records have a high degree of reliability. Identity clues from manual supplementation, interface notes, or non-critical auxiliary fields have a lower degree of reliability. The first step is to perform a format integrity check. This check examines whether the identity clues retain sufficient distinguishing information, such as whether the phone number is missing key digits, whether the address identifier is only retained at the regional level, whether the device identifier has truncated or contains abnormal characters, and whether the account identifier only retains a common prefix that cannot distinguish individuals. Next, a contextual consistency check is performed. This check examines whether the identity clues are consistent with the channel source, store affiliation, fulfillment status, and time location of the current consumption event. For example, if the same payment identifier corresponds to a continuous transaction chain within adjacent time periods, the consistency is high; if the address identifier is consistent with the delivery area retained in the fulfillment record, the consistency is high; if the device identifier crosses incompatible record sources within a short period, the consistency is low. After completing these three types of checks, an event-level identity description is generated for each consumption event. The meaning of event-level identity description is that it is the identity summary result formed after uniformly sorting out the credibility level, completeness level and context fit level of the consumption event on the five types of identity clues, which is used for subsequent event association filtering. The effect of this step is to transform the original identity clues into event-level identity descriptions with credible evidence, so that subsequent association analysis can distinguish between high credibility clues and low credibility clues, and improve the stability and transparency of the merging process.

[0090] Based on identity clue groups and event-level identity descriptions, candidate association filtering and association scoring are performed on consumption events in a unified consumption event sequence to generate a candidate association set. Specifically, candidate association filtering is performed on each pair of consumption events in the unified consumption event sequence. During filtering, stable clues in the identity clue group are compared first. The stable clues include telephone number identifiers, account identifiers, payment identifiers, and device identifiers that have passed source reliability and format integrity checks. When there are insufficient stable clues, address identifiers and context consistency results are introduced to assist in the judgment. The processing logic of candidate association filtering includes the following: if two consumption events share at least one type of highly reliable identity clue, they are included in the candidate set; if two consumption events do not have directly identical highly reliable clues, but there is a traceable association relationship in the account identifier, payment identifier, and device identifier, and the event context information is continuous, they are included in the candidate set; if two consumption events only share a single highly reliable identity clue, they are included in the candidate set. If the integer address fragments are similar and lack other supporting information, they will not be included in the candidate set. After the candidate screening is completed, the consumption event pairs that have entered the candidate set are subjected to association scoring. The association scoring does not use a single identical method for judgment, but assigns weights to five types of identity clues respectively. Telephone identifiers and account identifiers have higher weights because they have strong individual distinguishing ability. Payment identifiers and device identifiers have medium weights because there is a possibility of sharing or phased changes. Address identifiers have lower weights and are used to supplement and assist. When scoring, the consistency of each type of clue is first counted, and then the event-level identity descriptions of the two consumption events are combined to improve the consistency score of high-confidence clues and reduce the consistency score of low-confidence clues. Then, conflict information is checked. For example, if two events are obviously inconsistent in high-confidence account identifiers, or if there are incompatible payment identifiers in the same time period, the score is deducted. After screening and scoring, a candidate association set is formed. The candidate association set is a set that includes all consumption event pairs that meet the association conditions and their order of association strength. The effect of this step is to narrow down the potential association range in all consumption events to a candidate set with clear evidence, and to retain the association strength through hierarchical scoring, providing an interpretable basis for the subsequent generation of member merging groups.

[0091] The candidate association set is merged and aggregated, and conflict checking is performed to generate member merge groups. Index allocation is then performed on these member merge groups to generate a unified member index. Specifically, consumption event pairs with association scores higher than a preset threshold are selected from the candidate association set as the initial merging basis. The preset threshold is derived from previously set rules, determined based on identity clue weights, conflict deduction rules, and event context continuity requirements, and is used to exclude weakly associated event pairs from the merging starting point. Next, merge and aggregation are performed. Merge and aggregation refers to using high-scoring event pairs as the core and gradually absorbing other consumption events with stable associations to form several member merge groups. Consistency expansion rules are followed during absorption: if a new consumption event is consistent with most consumption events in the merge group in terms of high-credibility identity clues and has no obvious conflict in event context, it is included in that merge group; if a new consumption event is associated with multiple merge groups simultaneously, its overall association degree in each merge group is compared, and only the merge result with the highest association degree and no conflict is retained. Finally, conflict checking is performed. Conflict checking refers to checking again within the merge group for mutually exclusive high-credibility identity clues. The process involves identifying identity clues, incompatible timelines, and contradictory contextual trajectories. If a consumption event is found to have a stable conflict with the overall merge group, that consumption event is removed and re-entered into the merge judgment. After merging and conflict verification, a member merge group is formed. A member merge group is a set of consumption events belonging to the same member. Subsequently, index allocation is performed on each member merge group. During index allocation, a unique unified member index is assigned to each member merge group according to preset index generation rules. The preset index generation rules include uniqueness constraints, order constraints, and non-reusability constraints. Uniqueness constraints ensure that the same merge group corresponds to only one unified member index. Order constraints ensure that new indexes are added or expanded in a predetermined order. Non-reusability constraints ensure that previously allocated indexes are not reused. After index allocation, a unified member index is finally generated. The effect of this step is to further enhance the local correlation results at the consumption event level into stable merge results at the member level, and to establish a unified identity foundation for subsequent consumption trajectory rearrangement and preference deposition through the unified member index.

[0092] The present invention is further configured such that, in rearranging the unified consumption event sequence based on the unified member index and performing succession discrimination on the event context information, the following steps are included:

[0093] Based on the member merging groups corresponding to the unified member index, the unified consumption event sequence is subjected to attribution partitioning and temporal sorting to generate member event subsequences. Specifically, the unified consumption event sequence and unified member index generated in the previous steps are read. The unified member index comes from the identity clue association parsing results, and the member merging group comes from the consumption event set corresponding to the unified member index. During attribution partitioning, unified consumption events are read one by one, and their respective unified member indexes are used as attribution markers. All consumption events under the same unified member index are grouped into the same member merging group. After the attribution partitioning is completed, the consumption events within each member merging group are sorted temporally. The sorting is based primarily on the unified timestamp, which comes from the result of previous time alignment and offset correction and can reflect the unified temporal position of the consumption events. When two consumption events have the same unified timestamp or are within a very short time interval, the preset business order rules are invoked to perform concurrent resolution. The preset business order rules are derived from the event sequence relationship of the standard transaction chain, such as first forming an order, then forming a payment status, then forming a fulfillment status, and then forming a completion status, thereby determining the order of events within the same time neighborhood. After attribution division and time sequence sorting, a member event subsequence is generated. The meaning of a member event subsequence is a set of consumption events arranged in a unified time sequence under the same unified member index. The effect of this step is to compress all consumption events into a local time sequence chain for a single member, providing a clear event boundary and sequence basis for subsequent acceptance judgment, while avoiding misjudgment caused by different member events overlapping in time.

[0094] For adjacent consumption events in a member event subsequence, candidate construction and span filtering are performed to generate valid candidate event pairs for acceptance. Specifically, each member event subsequence is scanned sequentially in time, and the preceding and following consumption events are paired as adjacent consumption event pairs as initial acceptance candidates. Using an adjacent approach to construct candidates ensures that acceptance judgment focuses on consumption events that are temporally continuous and sequentially adjacent, guaranteeing that subsequent judgment results accurately reflect the natural extension of consumption intentions. After candidate construction, span filtering is performed on adjacent consumption event pairs. The span in the span filtering is derived from the uniform time interval between two consumption events, and is combined with the uniform event type and uniform product index of the preceding consumption events to determine the maximum allowable interval range for acceptance. The maximum interval range is given by a preset acceptance span table, which is derived from historical transaction chain statistics and business rule settings. Its function is to limit the allowable intervals between different event combinations. The time interval is considered; for example, payment events and verification events can have a longer interval, browsing events and order placement events can have a medium interval, and refund events after completion need to have their span determined in conjunction with refund rules; during the filtering process, if the actual time interval of adjacent consumption event pairs falls within the allowed range of this type of event combination, they are retained as valid candidate event pairs for acceptance; if they exceed the range, they are determined to be unable to be directly accepted and will not enter the subsequent context description process; the adjacent consumption event pairs used in this step are event pairs extracted from the member event subsequence in chronological order; the acceptance span table is a pre-set set of time sequence constraint parameters, which gives time distance thresholds according to the type of preceding event, the type of succeeding event, and the product relationship; the effect of this step is that by controlling the candidate range through local adjacency relationships and span constraints, the acceptance judgment is focused on event pairs with actual probability, improving the relevance and interpretability of the subsequent description results;

[0095] Extract the event context information of the preceding and subsequent consumption events from the valid candidate event pairs, perform time interval calculation, channel migration identification, promotion link comparison, and fulfillment status connection determination, and generate a context description. Specifically, extract the event context information of the preceding and subsequent consumption events for each valid candidate event pair. The event context information comes from the construction result of the preceding unified consumption event sequence and includes at least a unified timestamp, channel source information, store information, promotion information, and fulfillment association information. After extraction, first perform time interval calculation. Time interval calculation refers to subtracting the unified timestamp of the preceding consumption event from the unified timestamp of the subsequent consumption event to obtain the actual time interval between the two, and compare this time interval with the allowed span in step two to determine the degree of time continuity. Then, perform channel migration identification. Channel migration identification refers to determining whether there is a transferable migration path between the channel to which the preceding consumption event belongs and the channel to which the subsequent consumption event belongs. For example, when transferring from one source link to another source link, does it meet the preset transaction conversion rules? The preset transaction conversion rules come from the channel flow configuration table, which records the allowed channel transfer directions and transfer conditions. Then... The process involves comparing promotional links to determine if the promotional information of preceding and subsequent consumption events falls within the same promotional activity link, the same benefit usage link, or the same coupon link. During the comparison, it checks whether the activity identifiers are consistent, whether the coupon redemption identifiers have a sequential relationship, and whether the membership benefit usage status remains continuous. Following this, a fulfillment status continuity determination is performed. This determination, based on the fulfillment association information of the preceding and subsequent consumption events, determines whether there is a sequential continuity in their fulfillment processes. For example, if the preceding fulfillment status is "pending execution" and the subsequent fulfillment status is "executed," then the process is as follows: If the preceding sequence is completed but the following sequence returns to pending execution, then the connection is considered abnormal. After completing the above four processes, the continuity of time, channel migration results, promotion link comparison results, and fulfillment status connection results are integrated into a context description. The meaning of the context description is that it is the result of systematically depicting the connection environment of a pair of adjacent consumer events in four dimensions: time, channel, promotion, and fulfillment. The effect of this step is to transform whether the pair of adjacent events has a basis for connection into a contextual basis that can be explained item by item, so that the subsequent determination of the connection coefficient has a clear source and a transparent judgment path.

[0096] Extract the unified product index, unified event type, and store information of the preceding and following consumption events from valid candidate event pairs. Perform product correspondence comparison, event flow order determination, and store acceptance determination to generate a continuous semantic description of the events. Specifically, for each valid candidate event pair, further extract the unified product index, unified event type, and store information of the preceding and following consumption events. The unified product index comes from the product mapping result, the unified event type comes from the event normalization result, and the store information comes from the store field or fulfillment node field in the event context information. First, perform product correspondence comparison. The comparison, specifically the product correspondence comparison, involves checking whether there are any similarities, categories, specification conversions, packaging correspondences, or combination inclusions between the products corresponding to two consecutive consumption events. These relationships are derived from a product relationship table in a pre-defined standard product library. During the comparison, the first step is to determine if they belong to the same unified product index. If not, further checks are made to determine if they are at the same product level, belong to the same specification system, or are part of the same combination chain. Secondly, the event flow order is determined, which involves checking whether the unified event types of the preceding and subsequent consumption events conform to the same criteria. A preset event flow sequence table is created; this table, derived from a standard business event dictionary, records the allowed sequence of events. For example, browsing allows entry into order placement, order placement allows entry into payment, payment allows entry into fulfillment, fulfillment allows entry into completion, and completion allows for refunds within the rules' limits. Finally, a store acceptance determination is performed. This involves comparing the store information of two consecutive consumer events to determine if they belong to the same store, fulfillment node, fulfillment network, or have a traceable acceptance relationship. Store acceptance rules are derived from the store network configuration table and fulfillment node mapping table, and are used to define... The scope of acceptance between stores; after completing the product correspondence comparison, event flow sequence determination, and store acceptance determination, the three results are organized into a continuous event semantic description; the meaning of continuous event semantic description is to uniformly express the degree of continuity between two consecutive consumption events in terms of product semantics, event semantics, and store semantics; the effect of this step is to further advance the acceptance judgment from the external context to the internal event semantic level, so that the subsequent formation of cross-channel intention acceptance coefficient has comprehensive support of product consistency, event sequence consistency, and store acceptance consistency, thereby improving the accuracy and transparency of acceptance analysis.

[0097] The present invention is further configured such that the output of cross-channel intent reception coefficient and consumption trajectory fragment includes:

[0098] A joint determination is performed on the context description and the event semantic continuity description to generate the cross-channel intent acceptance coefficient corresponding to the valid acceptance candidate event pair. Specifically, the valid acceptance candidate event pair, the context description, and the event semantic continuity description generated in the previous steps are read first. The context description is derived from the results of sorting out the continuity of time, channel migration results, promotion link continuity, and fulfillment status of the preceding and following consumption events. The event semantic continuity description is derived from the results of sorting out the correspondence of the preceding and following consumption events in the unified product index, the unified event type flow order, and the store acceptance relationship. During the joint determination, the four types of factors in the context description are first classified. The time continuity... The continuity of events is categorized into high continuity, medium continuity, and low continuity based on the actual time interval within the allowed span; channel migration results are categorized into direct acceptance, conditional acceptance, and no acceptance based on preset transaction conversion rules; the continuity of promotional links is categorized into complete continuity, partial continuity, and no continuity based on the consistency of activity identifiers, coupon identifiers, and benefit identifiers; the continuity of fulfillment status is categorized into forward connection, weak connection, and conflict based on the sequential logic of fulfillment nodes and fulfillment stages; subsequently, the three types of factors in the semantic continuity description of events are graded; the continuity of product correspondence is determined according to the hierarchical order of the same product, similar products, specification conversion products, packaging corresponding products, and combined related products; event categories... The flow sequence is determined according to a preset event flow sequence table, categorized as sequentially valid, tolerable, or conflicting. Store acceptance relationships are differentiated based on whether the store is in the same location, part of the same fulfillment network, or has no acceptance relationship. After grading, a joint judgment is executed, employing a cumulative and conflict-deducting approach. First, basic scores are assigned to seven results: time continuity, channel migration, promotional link, fulfillment connection, product correspondence, event flow, and store acceptance. Then, scores are added to results with high reliability and continuity, deducted from conflicting results, and reduced for results with conditional acceptance. The basic scores and deduction rules are derived from a preset acceptance judgment parameter table, which is used in this embodiment. The system pre-sets and records the contribution order of various continuous and conflicting relationships in the consumer intent acceptance judgment. Then, it normalizes the accumulated results to a unified scale to generate a cross-channel intent acceptance coefficient. The cross-channel intent acceptance coefficient means that it is a quantitative result of whether a pair of valid acceptance candidate events belong to the same continuous expression of consumer intent under a unified member index. The higher the value, the stronger the continuity of consumer intent between the events. The effect of this step is to unify the judgment results scattered across multiple dimensions such as time, channel, promotion, fulfillment, product, event, and store into a single acceptance quantity, and retain all judgment sources, so that the subsequent segmentation process has a clear calculation basis and a transparent judgment path.

[0099] Based on the cross-channel intent carrying coefficient, the initial consumption trajectory fragments are generated by sequentially segmenting adjacent consumption events in the member event subsequence. Specifically, the member event subsequences corresponding to the same unified member index are processed one by one from front to back according to the sorting result, with the first consumption event as the starting point of the first fragment. Then, the cross-channel intent carrying coefficients corresponding to each pair of adjacent consumption events in the member event subsequence are read, and sequential segmentation is performed according to the preset segmentation threshold. The preset segmentation threshold comes from a pre-established segmentation judgment parameter table, which considers at least three factors: event type stability, product type change range, and member historical event rhythm. The event type stability comes from the continuity of the event types that have appeared in the member event subsequence; the higher the continuity, the lower the segmentation threshold. The product type change range comes from the corresponding adjacent consumption events. For product relationship levels, if they are still within the same product system or the same product relationship table, a lower continuity coefficient is allowed to maintain the same segment. If the product differences are large, the segmentation threshold is increased. The rhythm of member historical events is derived from the time distribution of historical consumption events corresponding to the unified member index. Members with smaller time interval fluctuations adopt a stricter segmentation standard, while members with larger time interval fluctuations adopt a more lenient segmentation standard. In specific processing, the cross-channel intent continuity coefficient of the current adjacent consumption event pair is compared with the current threshold. If the continuity coefficient reaches or exceeds the threshold, the subsequent consumption event is connected to the current segment. If the continuity coefficient is lower than the threshold, the segment is cut off at the subsequent consumption event, ending the current segment, and the subsequent consumption event is used as the starting point of the new segment. After processing the entire member event subsequence in sequence, the initial consumption trajectory segment is formed. The meaning of the initial consumption trajectory segment is that within the unified member index, the event continuity segment is obtained by first segmenting based on the strength of adjacent events. The effect of this step is to transform the linearly arranged consumption events into multiple initial segments with local consumption intention continuity, and to incorporate the differences in consumption rhythm of different members into the segmentation process through the threshold self-adaptation mechanism, thereby improving the rationality of the initial trajectory segmentation.

[0100] The initial consumption trajectory segments are processed by performing internal segment continuity checks and adjacent segment boundary adjustments to generate consumption trajectory segments. Specifically, internal segment continuity checks are performed on each initial consumption trajectory segment. The objects of internal segment continuity checks are the cross-channel intent continuity coefficients and corresponding judgment sources of all adjacent consumption event pairs within the segment. During the check, each item is checked to see if there are consecutive low continuity segments, event flow conflicts, sudden changes in product relationships, and reversals in fulfillment status. Consecutive low continuity segments refer to the intervals within the same initial consumption trajectory segment where the continuity coefficients of multiple adjacent event pairs are consistently lower than the segment's stable threshold. Event flow conflicts refer to the occurrence of events of the same type that violate the preset event flow order table. Sudden changes in product relationships refer to the absence of a continuity relationship between adjacent event corresponding products in the product relationship table. Reversals in fulfillment status refer to the subsequent fulfillment status reverting to a state before the previous fulfillment status. If the internal segment continuity check finds the above anomalies, the initial consumption trajectory is adjusted according to the position of the first occurrence of the anomaly. The segments are further divided into more stable internal sub-segments. After completing the internal continuity verification of the segments, boundary correction is performed on adjacent segments. The boundary correction targets the boundary events of two adjacent initial consumption trajectory segments and their neighboring events. During correction, the cross-channel intent continuity coefficient at the boundary of two adjacent segments is compared first, and then the continuity of the product relationship, event flow relationship, and store continuity relationship on both sides of the boundary is checked. If the boundary continuity coefficient is higher than the boundary merging threshold, and the continuity conditions are met on both sides of the boundary in the three dimensions of product, event, and store, then the two adjacent segments are merged into one segment. If the boundary continuity coefficient is in the middle range, the continuity of the promotion link and the completion status of the fulfillment status are further considered to determine whether the boundary position needs to be adjusted, such as moving one event forward or one event backward. If the boundary continuity coefficient is low or there is a conflict on both sides of the boundary, the boundary remains unchanged. The boundary merging threshold and the boundary adjustment threshold are derived from the preset boundary correction parameter table, which is set based on the historical continuity link statistics. After completing the internal continuity verification of segments and the boundary correction of adjacent segments, the final consumption trajectory segment is generated. The meaning of a consumption trajectory segment is, under the unified member index, an ordered set of consumption events with stable continuity, coordinated event sequence, continuous product relationship and corrected boundary position. The effect of this step is to perform stability verification and boundary optimization on the initial segmentation results, making the output consumption trajectory segment more suitable as input for subsequent preference deposition, and making the source of segment boundaries clear, the adjustment rules clear, and the results traceable.

[0101] The present invention is further configured such that the step of performing preference deposition on consumption trajectory segments and generating dynamic member preferences by combining cross-channel intent reception coefficients includes:

[0102] Based on the consumption events in the consumption trajectory segments, a unified product index, unified event type, and event context information are extracted, and preference semantic mapping is performed to generate event preference primitives. Specifically, the consumption trajectory segments corresponding to the unified member index are read. Each consumption trajectory segment consists of a set of consumption events ordered after continuation discrimination and boundary correction. For each consumption event in the segment, a unified product index, unified event type, and event context information are extracted. The unified product index comes from the standard product library mapping results and can indicate the position of the product corresponding to the consumption event in the unified product system. The unified event type comes from the event normalization results and can indicate whether the consumption event belongs to browsing, ordering, payment, verification, refund, fulfillment completion, etc. A standard behavior is defined as follows: the event context information is derived from the preceding unified consumption event construction steps, including at least channel source information, store information, promotion information, fulfillment information, and unified timestamp; after extraction, preference semantic mapping is performed; preference semantic mapping refers to transforming the information from the above three sources into unified preference expression units that can be used for subsequent preference deposition; in specific processing, firstly, the category level, brand level, specification level, and price level information in the standard product library are queried based on the unified product index, then the behavior intensity level and behavior direction attribute in the preset behavior semantic table are queried based on the unified event type, and then the channel participation method, promotion participation status, and fulfillment stage information in the event context information are organized into context labels. Then, according to the preset preference mapping rules, product hierarchical features, behavioral semantic features, and context labels are mapped to the same preference description structure to form event preference primitives. The meaning of an event preference primitive is that a certain consumption event is the smallest semantic unit that contributes to the formation of a member's preference, and it contains at least three parts: product interest orientation, behavioral evidence strength, and context correction factors. The effect of this step is to transform the original consumption event into a preference base object that can be directly accumulated, compared, and integrated in the future, so that consumption behaviors from different sources, different event types, and different product hierarchies have a unified preference semantics.

[0103] Combining cross-channel intent acceptance coefficients, unified event types, and time information, the event deposition intensity is used to determine the deposition weights. Specifically, the deposition weights are determined for each consumption event within each consumption trajectory segment. The sources of deposition weights include cross-channel intent acceptance coefficients, unified event types, and time information. Cross-channel intent acceptance coefficients are derived from the preceding acceptance discrimination step, representing the continuity of adjacent consumption events in terms of consumption intent. Unified event types are derived from event normalization results, used to distinguish the strength of different behaviors' contributions to preference formation. Time information is derived from unified timestamps, used to measure the distance of consumption events from the current analysis point. During processing, the cross-channel intent acceptance coefficients within the consumption trajectory segment are first partially adjusted. Partial adjustment refers to combining the acceptance coefficients between a consumption event and its preceding and following adjacent consumption events to form the local acceptance support level of that consumption event. For consumption events located at the beginning of a segment, the acceptance coefficient between them and subsequent events is prioritized. For consumption events located in the middle of a segment, the acceptance coefficients in both the preceding and following directions are considered. For consumption events at the end of a segment, the continuity coefficient between the event and its preceding events is used first. Then, a pre-defined event role table is consulted based on the unified event type. This table, derived from business rules, illustrates the strength of evidence for different event types in preference formation. For example, payment completion and verification completion correspond to higher weights, browsing and promotional triggers to medium weights, and refunds and cancellations to negative correction weights. Next, a unified timestamp is read, and the time decay level is calculated based on the current analysis point. The time decay level reflects the recentity of the event; the closer the time is to the current analysis point, the greater its impact on recent preferences. Finally, the local continuity support level, event type weight, and time decay level are combined to generate a deposition weight. The deposition weight represents the share of a consumption event in the current consumption trajectory segment when it enters preference accumulation. The effect of this step is to integrate the continuity of consumption intention, the strength of behavioral evidence, and the recentity of time into the event-level weight formation process, ensuring that preference deposition is based on a sourced, differentiated, and timely contribution assessment.

[0104] The process involves sequentially depositing event preference primitives and deposition weights to generate segment preference representations, and organizing the cross-channel intent acceptance coefficients within consumption trajectory segments to generate segment acceptance stability information. Specifically, for each consumption trajectory segment, preference deposition is performed according to the chronological order of consumption events within the segment. During processing, an initial preference state for the segment is first established. This initial preference state originates from a preset initial configuration and is represented by an empty or zero state, indicating that no valid preference information has been accumulated at the beginning of the segment. Then, the event preference primitives and deposition weights for each consumption event are read sequentially. For each consumption event, the segment state is updated by combining the retained portion of the existing segment state with the absorbed portion of the current event preference primitive. The proportion of the retained portion is determined by the supplementary portion of the deposition weights, and the proportion of the absorbed portion is determined by the deposition weights themselves. After this processing, consumption events with higher deposition weights have a stronger correction effect on the current segment preference state, while consumption events with lower deposition weights have a weaker disturbance effect on the existing preferences of the segment. After updating each event, the result is obtained at the end of the segment. Segment preference representation; the meaning of segment preference representation is the comprehensive expression of a consumption trajectory segment in terms of product interest, behavioral tendency, and contextual preference; at the same time, all cross-channel intention reception coefficients within the segment are processed to generate segment reception stability information; during processing, the overall level of each reception coefficient within the segment is first statistically analyzed, and then the existence of continuous low reception segments, local conflict segments, and abrupt change segments is checked; if the vast majority of reception coefficients within the segment remain at a high level, the segment reception stability is considered high; if there are multiple continuous low reception positions, the stability is reduced; if the low reception position only exists locally but the events before and after it are continuous in terms of product and event type, appropriate adjustments are made; after processing, segment reception stability information is formed; the meaning of segment reception stability information is the summary result of the overall state of intention continuity within the consumption trajectory segment; the effect of this step is to accumulate event-level preference contributions sequentially into segment-level preference results and simultaneously output the degree of reception stability within the segment, providing a reliable basis for subsequent cross-segment fusion;

[0105] Based on segment preference representations, segment continuity stability information, and time information, cross-segment fusion is performed to generate long-term and recent preferences. Specifically, cross-segment fusion is performed on all consumption trajectory segments formed under the same unified member index. Each segment corresponds to a segment preference representation and segment continuity stability information, as well as a segment end time. The segment end time is derived from the unified timestamp of the last consumption event in that segment. First, long-term preferences are generated, which represent the stable product interests and behavioral tendencies of members over a longer period. During processing, all segment preference representations are assigned long-term fusion weights. The long-term fusion weights are jointly determined by segment continuity stability information, the cumulative deposition level within the segment, and the distance between the segment end time and the current analysis point. Segments with high continuity stability information have higher long-term reference value. Segments with high cumulative deposition levels within the segment indicate that they contain more valid preference evidence. Although segments with earlier end times will be subject to some time decay, if their stability and cumulative deposition levels are high, they still retain a high long-term fusion share. All segments are aggregated according to the long-term fusion weights to form long-term preferences. Subsequently, recent preferences are generated, which represent the current trend of member interest changes. During processing, the recent fusion weights place greater emphasis on the recentity of the segment's end time while retaining checks on the segment's continuity information. Segments whose end time is closer to the current analysis point and whose continuity information is higher have a higher proportion in recent preferences. After recent fusion, recent preferences are formed. The meanings of long-term preferences and recent preferences are respectively: stable preference results reflecting members' continuous consumption habits and recent preference results reflecting members' current stage of interest tendencies. The effect of this step is to organize the scattered segment preference results into two preference results with different time scales, so that subsequent dynamic fusion can both retain long-term patterns and reflect recent changes.

[0106] Dynamic fusion of long-term and recent preferences is performed to generate dynamic member preferences. Specifically, after reading long-term and recent preferences, the degree of preference difference between them is first calculated. The degree of preference difference is derived from a step-by-step comparison of long-term and recent preferences in a unified preference description structure. The comparison includes at least product interest direction, behavioral tendency direction, and contextual preference direction. If long-term and recent preferences are similar in most dimensions, the member's overall preferences are considered relatively stable. If there are significant differences between the two in multiple core dimensions, it is considered that preference drift has occurred recently. Then, the overall segment stability level is calculated. The overall segment stability level is derived from the summary result of segment continuity stability information of all consumption trajectory segments under the unified member index, which is used to reflect whether recent preference changes are based on stable segments. Finally, a dynamic fusion ratio is generated according to a preset fusion rule. The preset fusion rule is derived from the dynamic preference fusion parameter table. The rules stipulate that when the difference between long-term and short-term preferences is small and the overall segment stability level is high, the dynamic fusion ratio is tilted towards long-term preferences; when the difference between short-term and long-term preferences is large and the short-term related segments are relatively stable, the dynamic fusion ratio is tilted towards short-term preferences; when the difference between short-term preferences is large but the segment stability level is insufficient, the dynamic fusion ratio remains conservative to avoid directly absorbing short-term noise into the final preference; after determining the fusion ratio, long-term and short-term preferences are combined proportionally to generate dynamic member preferences; the meaning of dynamic member preferences is the current member preference result formed by absorbing recent effective preference changes while retaining historical stable preferences; the effect of this step is that the final preference result can simultaneously reflect long-term consumption patterns and recent behavioral changes, and the fusion process has clear sources, clear conditions, and clear weighting basis, providing stable and updatable preference input for subsequent product recommendations.

[0107] The present invention is further configured such that the step of collecting product operation records, normalizing and associating the execution status of product operation records to generate product constraint information includes:

[0108] Collect product operation records and extract product identifiers, specification information, channel operation identifiers, store operation identifiers, operation time, inventory information, price information, promotion execution information, fulfillment execution information, and source information to generate product operation atoms. Specifically, first access business records directly related to the product operation status. These product operation records are at least derived from inventory records, price records, promotion records, and fulfillment records. Product identifiers are derived from product codes, product primary keys, or standard material numbers to identify the product entity. Specification information is derived from capacity, weight, quantity, packaging dimensions, or specification descriptions to distinguish differences between similar products at different specification levels. Channel operation identifiers are derived from sales channels, interface sources, or business channel codes to indicate the channel dimension corresponding to the record. Store operation identifiers are derived from store numbers, fulfillment node numbers, or operation unit numbers to indicate the store dimension corresponding to the record. Operation time is derived from price effective time, inventory update time, promotion execution time, or fulfillment status update time to indicate the time position of the record. Inventory information is derived from the ledger database. Information on inventory, pre-held inventory, frozen inventory, safety stock, and in-transit replenishment; price information from current execution price, reference price, listed price, and price adjustment records; promotion execution information from activity identifiers, discount rules, promotion effective status, and promotion ended status; fulfillment execution information from deliverable status, self-pickup status, verification available status, and fulfillment completed status; source information from original system tags, interface tags, and update link tags; after extraction, it is encapsulated into product operation atoms according to a preset field structure; the meaning of a product operation atom is that, in a certain source record, it is the smallest data unit that structures the operational status of a product in a specific channel, a specific store, and a specific time location; the preset field structure also records whether the field is missing, the field update time, and the field source priority, so that subsequent status unification has a clear basis; the effect of this step is to unify the product operation information that was originally scattered in different systems, different granularities, and different naming methods into the same data structure, providing a unified input basis for subsequent product index mapping, time slicing, and status fusion;

[0109] Product identification and specification information are mapped to a unified product index space. Time window slicing is performed on the operational time. Based on the unified product index, channel operation identifier, store operation identifier, and slice time, product operation atoms are aggregated to generate scenario product units. Specifically, a unified mapping is first performed on product identification and specification information. The unified product index space comes from a pre-established standard product library, which at least stores the product master index, specification hierarchy, packaging hierarchy, and combination hierarchy. During mapping, the product identifier is used as the first-level matching condition. If different source systems use different product codes, then specification information is combined for a second-level verification. Specification information verification includes checking whether the specification values, specification units, packaging quantities, and description text fall into the same specification template are consistent. After the mapping is completed, each product operation atom obtains a unique unified product index. Then, time window slicing is performed on the operational time. The meaning of time window slicing is to divide the continuous operational time into preset time windows. The data is divided into discrete time periods, allowing for comparison of inventory, pricing, promotions, and fulfillment information within the same time frame at a unified time granularity. The preset time window is derived from a combination of operational status update frequency and recommendation processing frequency; for example, a shorter window is used when the update frequency is high, and a longer window is used when the update frequency is low. After slicing the time window, the unified product index, channel operation identifier, store operation identifier, and slice time are combined into a scenario key. Product operation atoms with the same scenario key are then grouped into the same set to form scenario product units. A scenario product unit is a scenario-level data unit resulting from the aggregation of multiple product operation atoms under the conditions of unified product, channel, store, and time window. The effect of this step is to elevate product operation records from the original source granularity to a unified scenario granularity, enabling subsequent status normalization and correlation organization to be performed under the same comparison benchmark, avoiding the distortion of constraint information caused by the direct mixing of data from different time locations, store locations, and channel locations.

[0110] The execution status of inventory, price, promotion, and fulfillment information within a scenario's product unit is unified, and combined with source information to generate a scenario status quantity. Specifically, inventory, price, promotion, and fulfillment status are processed separately for each scenario's product unit. When unifying inventory status, book inventory, pre-held inventory, frozen inventory, safety stock, and in-transit replenishment information are first organized into a sellable inventory layer, a restricted inventory layer, and a pending replenishment inventory layer. Then, a unified inventory status is formed based on preset inventory conversion rules. These preset inventory conversion rules are derived from the inventory business rules table, which stipulates that frozen inventory cannot be directly considered sellable inventory, in-transit replenishment is converted into partial sellable reserves based on arrival certainty, and safety stock is a reserve quantity that cannot be directly used. When unifying price status, the current execution price, reference price, and price adjustment status are first read, and then the price is determined according to the price effectiveness order and price... Price priority is used to form a unified price status. Price priority is derived from the price strategy table, typically placing the currently effective execution price with the highest priority, and historically invalidated prices and prices awaiting activation with lower priority. When unifying promotion status, the activity identifier, discount rules, effective period, and execution status are read first, and then the promotion execution information is organized into standard statuses such as promotion effective, promotion awaiting activation, promotion invalidated, and promotion conflict. When unifying fulfillment status, fulfillment execution information such as deliverable, self-pickup available, redeemable, fulfillment closed, and fulfillment completed is read first, and then organized into a unified fulfillment status according to the order of fulfillment stages. After the above four types of states are unified, they are merged and integrated based on the source information. During the merging and integration, source priorities are first set for different sources. The source priorities are derived from data quality management rules, with direct business system sources taking precedence over intermediate cache sources, and real-time interface sources taking precedence over offline synchronization sources. Then, the update times of multiple records within the same product unit in the same scenario are compared, with newer records taking precedence over older records. If different sources conflict on the same state, the retained value is determined in the order of higher source priority, more recent update time, and higher field completeness. Field completeness is determined by checking whether the inventory field, price field, promotion field, and fulfillment field are complete. After the merging and integration is completed, a scenario state quantity is formed. The meaning of the scenario state quantity is that the inventory state, price state, promotion state, and fulfillment state of a certain product in a certain channel, a certain store, and a certain time window are uniformly organized. The effect of this step is that it converges the product operation information from multiple sources, multiple update times, and multiple business states into a scenario state that can be directly used for subsequent filtering, and the fusion rules are clear, the sources are traceable, and the state generation path is clear.

[0111] Based on the specification relationships, packaging relationships, combination relationships, and fulfillment connection relationships corresponding to the unified product index, the scenario product units are associated and organized to generate associated descriptive quantities. Specifically, the product relationship information corresponding to the unified product index is read first. The specification relationship comes from the specification hierarchy table in the standard product library, which is used to identify the conversion relationship of the same product under different quantities, capacities, or weights. The packaging relationship comes from the packaging hierarchy table, which is used to identify the subordinate relationship between single items and packaged items. The combination relationship comes from the combination relationship table, which is used to identify the composition relationship between products and combination packages, sets, or combination sales units. The fulfillment connection relationship comes from the fulfillment network mapping table, which is used to identify the acceptable relationship between different product forms or different operating units at fulfillment nodes. Then, the association is organized for each scenario product unit. During the association organization, it is first checked whether there is a specification corresponding item in the current unified product index. If there is, the corresponding specification unit, conversion order, and substitutable level are recorded. Then, it is checked whether there is a packaging corresponding item. If there is, the packaging inclusion relationship and packaging split relationship are recorded. Then, it is checked whether there is a packaging corresponding item. If a combination of corresponding items exists, the combination members and constraints are recorded. Finally, the existence of fulfillment connection relationships is checked; if so, the connectable stores, connectable channels, and connectable fulfillment methods are recorded. After completing the four types of relationship checks, they are uniformly organized into association description quantities. The meaning of association description quantities is the structured expression of the business relationships between the current scenario's product unit and other related product units at the levels of specifications, packaging, combination, and fulfillment. Association description quantities not only retain whether a relationship exists, but also the direction, level, and priority of the relationship. For example, in specification relationships, it records whether the conversion is from small to large specifications or from large to small specifications; in packaging relationships, it records whether the packaging contains individual items or individual items are aggregated into packaging; in fulfillment connection relationships, it records whether the delivery path is complete. The effect of this step is to supplement cross-product constraints that cannot be directly expressed by a unified product index into the product-side description, enabling subsequent recommendation filtering to identify alternative items, packaging items, combination items, and fulfillment delivery items, improving the completeness of product constraint information and business consistency.

[0112] The unified product index, channel operation identifier, store operation identifier, slice time, scene status quantity, and associated description quantity are encapsulated to generate product constraint information. Specifically, after completing state normalization and association organization for each scene product unit, the unified product index, channel operation identifier, store operation identifier, slice time, scene status quantity, and associated description quantity are uniformly encapsulated to form product constraint information. During encapsulation, basic identifier information, including the unified product index, channel operation identifier, store operation identifier, and slice time, is written first in a fixed field order to indicate the object scope corresponding to the product constraint information. Then, the scene status quantity is written to indicate the current product's inventory, price, promotion, and fulfillment status within the corresponding scope. Finally, the associated description quantity is written to indicate the external links of the product at the specifications, packaging, combination, and fulfillment levels. The process involves adding a source traceability marker and an update time marker to allow for tracing back the constraint source in case of conflicts during the recommendation phase. The generated product constraint information can be understood as a constraint object that uniformly expresses the sellability, priceability, promotionability, fulfillability, and related substitutability of a product within a specific operational scope, oriented towards the recommendation and screening steps. This constraint object is in the same processing chain as the previously formed consumption trajectory fragments and member dynamic preferences; the former provides the product-side boundary, and the latter provides the member-side demand. The effect of this step is to ultimately form product constraint information that is structurally complete, clearly sourced, consistently in state, and clearly related, enabling subsequent related screening steps to directly call upon it without having to trace back the original product operation records again, thereby ensuring that the expression of product constraints in the recommendation chain is sufficient, transparent, and traceable.

[0113] The present invention is further configured such that the association filtering based on consumption trajectory fragments and product constraint information includes:

[0114] Extract the unified product index, unified event type, and event context information from the consumption trajectory segments, and match them with the unified product index, channel operation identifier, store operation identifier, and slice time from the product constraint information to generate candidate product items. Specifically, read the consumption trajectory segments corresponding to the unified member index. The consumption trajectory segments are derived from the results of the preceding inheritance discrimination and boundary correction, representing the event sequence formed by the same member under a continuous consumption intention. For each consumption trajectory segment, extract the unified product index, unified event type, and event context information carried by the consumption event within the segment. The unified product index is derived from the standard product library mapping results and is used to identify the standard product corresponding to the consumption event. The unified event type is derived from the event normalization results and is used to identify the behavioral attributes of the consumption event in the business chain. The event context information is derived from the unified consumption event sequence construction steps and includes at least channel source information, store information, promotion information, fulfillment information, and unified timestamp. At the same time, read the set of product constraint information. Product constraint information originates from the status normalization and correlation organization results of product operation records, and includes at least a unified product index, channel operation identifier, store operation identifier, slice time, scenario status quantity, and associated description quantity. During matching, the unified product index is first used as the primary matching key to filter out product constraint information directly consistent with historical products in the consumption trajectory segment. Next, channel source information and channel operation identifiers are used for channel-level comparison to confirm whether the consumer-side channel and the product-side operation channel are located in the same channel system or within a permitted channel system. Finally, store information and store operation identifiers are used for store-level comparison to confirm whether the consumer-side store and the product-side store are consistent, located in the same fulfillment network, or whether there is a permitted fulfillment network. The process involves mapping store relationships; finally, comparing the consumption time location corresponding to a unified timestamp with the slice time in a time dimension, retaining valid product constraint information within a preset time window; the preset time window is derived from the collaborative setting of the recommendation execution cycle and the product operation status update cycle, used to ensure the availability of matching results in terms of time; after completing the above multi-dimensional matching, candidate product items are formed. The meaning of candidate product items is that they are product constraint units that have a direct correspondence with consumption trajectory segments in terms of product identification, channel affiliation, store affiliation, and time location; the effect of this step is to establish a clear mapping between the historical consumption expression in the member's consumption trajectory and the current operation status on the product side, so that subsequent screening is based on a traceable correspondence;

[0115] Based on the inventory status, price status, promotion status, and fulfillment status in the scenario state variables, an initial candidate product set is generated by performing constraint filtering on candidate product items. Specifically, the generated candidate product items are read, and their inventory status, price status, promotion status, and fulfillment status are checked one by one in the scenario state variables. The scenario state variables are derived from the merged and integrated results of product operation records under a unified product index, unified channel, unified store, and unified time window. The inventory status represents the current product's available inventory level, restricted inventory level, and replenishment availability level within the corresponding operating scope. The price status represents the current execution price, price effectiveness status, and price validity range. The promotion status represents whether the promotion rules are effective, whether promotion resources are available, and whether promotion restrictions are met. The fulfillment status represents whether the product is available for delivery, self-pickup, redemption, or is in a closed state within the current operating scope. During constraint filtering, an inventory constraint check is performed first. The inventory constraint check is based on whether the available inventory has reached a preset available inventory threshold, while also referring to frozen inventory and safety stock limits, eliminating candidate product items that cannot form actual supply. The preset available inventory threshold is derived from the inventory management parameter table and is used to prevent products with inventory close to the lower limit from entering the recommended set. Then, a price constraint check is performed. Price constraint checks exclude candidate product items with invalid prices, pending prices, or price conflicts, while retaining product items with stable prices that can be directly applied. Next, promotion constraint checks are performed, reading the activity activation flag, benefit matching flag, and discount rule restrictions in the promotion status to eliminate candidate product items with invalid promotion status, unmet restrictions, or conflicting promotional resources. Following this, fulfillment constraint checks are performed, reading the delivery availability, self-pickup availability, and verification availability information in the fulfillment status to eliminate candidate product items with fulfillment closed, interrupted fulfillment paths, or unacceptable stores. After completing these four constraint checks, candidate product items that simultaneously meet inventory, price, promotion, and fulfillment requirements are retained, forming an initial candidate product set. This initial candidate product set refers to the set of products that already have corresponding product identifiers and can enter the subsequent recommendation expansion process under the current operational status. The effect of this step is to further compress candidate product items that only meet basic matching relationships into an initial set with real operational availability, thus ensuring that subsequent expansion and ranking are based on products that are salable, priced, promoted, and fulfillable.

[0116] Based on the specification relationships, packaging relationships, combination relationships, and fulfillment connection relationships in the associated descriptive values, the initial candidate product set is expanded and deduplicated to generate a new candidate product set. Specifically, the associated descriptive values ​​corresponding to each product item in the initial candidate product set are read. These associated descriptive values ​​originate from the association processing results in the product constraint information formation step, recording the acceptability relationships between the current product and other products at the specification, packaging, combination, and fulfillment network levels. Specification relationships are used to identify the conversion and acceptance order of the same product across different specifications; packaging relationships are used to identify the inclusion and separation relationships between individual items and their packaging; and combination relationships are used to identify the current... Whether the product belongs to a certain sales unit, or whether a certain sales unit consists of the current product; fulfillment connection relationship is used to identify whether the current product has a feasible path in other stores, other fulfillment nodes, or other fulfillment methods; when performing association expansion, first expand the candidate products based on specification relationship to have the same product system as the initial candidate products but different specifications, and mark them according to specification conversion direction and substitution level; then expand the candidate products based on packaging relationship to have packaging subordinate or packaging split relationship with the current product, and record the packaging level priority order; subsequently expand the candidate products based on combination relationship to have combination inclusion or combination member relationship with the current product, and classify the combination. Completeness is verified, including whether all bundled members are present, whether bundled promotions are effective, and whether bundled fulfillment is closed. Finally, based on fulfillment connection relationships, candidate items that can be fulfilled at other fulfillment nodes with the current product are expanded, and their fulfillment connection paths are recorded. After completing the association expansion, deduplication is performed. During deduplication, identical product items are first merged according to a unified product index. For product items with different specifications but convertible to the same recommended object, only the one with the higher priority is retained according to a preset priority. For product items repeatedly introduced by packaging and bundled relationships, the retention path is determined based on the association source priority table. For duplicate product items expanded from fulfillment connection relationships, the retention path is determined based on the fulfillment connection relationship. The optimal items are retained for order path completeness, store acceptance consistency, and time validity. The preset priority table and order path completeness rules are derived from the product relationship configuration table and the order network configuration table. After deduplication, a candidate product set is finally generated. The candidate product set means that it not only meets the consumption trajectory matching and product operation constraints, but also expands and supplements the specifications, packaging, combination, and order relationship, and is formed after deduplication. The effect of this step is to ensure that the recommended candidate range retains directly matching products and includes related products with clear business acceptance relationships. At the same time, the deduplication rules eliminate duplication and conflict, ensuring that the candidate set has a clear structure, clear source, and consistent constraints.

[0117] The present invention is further configured such that the step of generating product recommendation results by combining member dynamic preferences includes:

[0118] The process involves performing preference matching between member dynamic preferences and the candidate product set to generate preference matching results for the candidate products. Specifically, it reads the member dynamic preferences and candidate product set generated in the previous steps. The member dynamic preferences are derived from the result of preference deposition on consumption trajectory segments, and contain at least long-term and recent preference components. They are stored in a unified preference structure to preserve the product interest direction, behavioral preference direction, and contextual preference direction. The candidate product set is derived from the result of association filtering between consumption trajectory segments and product constraint information. Each candidate product in the set carries a unified product index, scenario state quantity, and associated description quantity. When performing preference matching, the process first queries the category level, brand level, specification level, price level, and combination level in the standard product library based on the unified product index. Then, it compares the product information with the corresponding preference items in the member dynamic preferences item by item. The comparison process includes at least the judgment of product category fit, brand stability, specification adaptability, price level consistency, and promotion response adaptability. Product category fit is used to determine whether the category of the candidate product falls within the interest range covered by both the member's long-term and recent preferences; brand stability is used to determine whether the candidate product is consistent with the member's established brand preferences; specification fit is used to determine whether the candidate product's specification level is consistent with the member's historical consumption structure; price level consistency is used to determine whether the current price position of the candidate product falls within the member's acceptable price range; promotion response fit is used to determine whether the promotion status of the candidate product matches the member's promotion sensitivity in dynamic preferences. After the above comparisons are completed, preference matching results are generated according to preset preference matching rules. The preset preference matching rules are derived from the preference matching parameter table, which clearly specifies the priority order and judgment level of the five factors of product category, brand, specification, price, and promotion in the matching process. The meaning of the preference matching result is to provide a structured expression of the degree of fit between each candidate product and the current member's dynamic preferences, including at least the primary matching level, secondary matching level, and mismatch reason marker. The effect of this step is to establish a one-to-one matching relationship between member preference information and product candidate information, providing a clear preference basis for subsequent trajectory support determination and recommendation ranking.

[0119] Combining consumption trajectory fragments and cross-channel intent acceptance coefficients, trajectory support determination is performed on the candidate product set, generating trajectory support results corresponding to the candidate products. Specifically, the consumption trajectory fragments corresponding to the current member, as well as the cross-channel intent acceptance coefficients within and at the fragment boundaries, are read. The consumption trajectory fragments are derived from the event sequence segmentation results under the unified member index, representing the continuous consumption intent formed by the member at different stages. The cross-channel intent acceptance coefficients are derived from the acceptance discrimination results, representing the degree of continuity of adjacent events in the same consumption intent chain. When determining the trajectory support, the candidate products are first checked for correspondence with historical products in the consumption trajectory fragments. The correspondence checks include direct product correspondence checks, specification relationship correspondence checks, packaging relationship correspondence checks, combination relationship correspondence checks, and fulfillment connection correspondence checks. The direct product correspondence check is used to determine whether the candidate product is a standard product that has appeared in the trajectory fragment. The specification relationship correspondence check is used to determine whether the candidate product belongs to the same acceptable item in the same specification system as the historical product. The packaging relationship correspondence check is used to determine whether the candidate product has a packaging subordination relationship with the historical product. The combination relationship correspondence check is used to determine whether the candidate product is in the same combination chain as the historical product. The fulfillment alignment check is used to determine whether a candidate product is within the acceptable range of the fulfillment network of the historical consumption event. After completing the product alignment check, the position and acceptance strength of the corresponding product in the consumption trajectory segment are checked. If the historical product corresponding to a candidate product is located in a consumption trajectory segment with high acceptance stability, and the cross-channel intention acceptance coefficient of its surrounding events remains at a high level, then the candidate product is considered to have strong trajectory support. If the candidate product is only related to isolated products in low-stability segments, or the correspondence is only at the weak specification substitution level, then the trajectory support is considered weak. If the candidate product has no... If a direct or indirect connection is found in any consumption trajectory segment, the trajectory support is deemed to be missing. To ensure transparency, the source of support is recorded in the trajectory support results, including the corresponding segment number, the corresponding product relationship type, the segment connection stability level, and the degree of connection continuity. The meaning of the trajectory support results is a structured expression of whether each candidate product is supported by the member's existing consumption trajectory and the strength of that support. The effect of this step is to establish a connection between candidate products and the member's historical continuous consumption intentions, so that the recommendation process not only relies on static preference matching but also combines the connection evidence in the real consumption chain.

[0120] Based on preference matching results and trajectory support results, the candidate product set is recommended and ranked to generate product recommendation results. Specifically, the preference matching results and trajectory support results are first merged. During merging, a recommendation judgment unit is established for each candidate product. The recommendation judgment unit includes at least the preference matching level, the reason for preference mismatch, the trajectory support level, the source of trajectory support, and the current constraint status of the product. Then, recommendation determination is performed, which adopts a hierarchical screening rule. The first layer of screening checks whether the candidate products simultaneously meet the basic requirements of preference matching and trajectory support. The basic requirements of preference matching are derived from a preset matching threshold table and are used to exclude products that deviate significantly from the dynamic preferences of members. The basic requirements of trajectory support are derived from a trajectory support rule table and are used to exclude products that lack effective support in the consumption trajectory. The second layer of screening classifies the candidate products that pass the first layer of screening into recommendation levels. The recommendation level classification comprehensively considers the product category fit, brand stability, specification adaptability, price level consistency, and promotion response adaptability in the preference matching results, as well as the segment stability, product corresponding level, and trajectory support results. The process proceeds according to the degree of continuity. If a candidate product is at a high level in both preference matching and trajectory support, it enters the priority recommendation layer. If only one is at a high level and the other is at a medium level, it enters the suboptimal recommendation layer. If both are at a low level, it does not enter the final recommendation result. After the recommendation is determined, the results are sorted. When sorting the results, they are first sorted by preference matching priority within the same recommendation layer, and then by trajectory support stability. When two items are still similar, they are then sorted by inventory availability, price effectiveness, and fulfillment availability in the product constraint information. Inventory availability, price effectiveness, and fulfillment availability are derived from the scenario state variables in the product constraint information to ensure that the final recommendation result is executable on the product side. After the above determination and sorting, the product recommendation result is generated. The product recommendation result means that, for the current member at the current processing point, the recommended products are an ordered set after being jointly filtered by preference matching, trajectory support, and product constraints. The effect of this step is to unify preference information, trajectory evidence, and product constraints into the same recommendation decision chain, and output recommendation results with clear sources and clear sorting criteria.

[0121] Based on the feedback records corresponding to the product recommendation results, update the cross-channel intent reception coefficient and member dynamic preferences. Specifically, continuously collect feedback records generated after the product recommendation results are output. These feedback records originate from exposure records, click records, add-to-cart records, order records, redemption records, refund records, and ignore records corresponding to the recommendation results. After collection, first align the feedback records with the corresponding recommended products, recommendation time, and recommendation location, then establish feedback associations with consumption trajectory segments and cross-channel intent reception coefficients. When establishing feedback associations, check which recommended product triggered a particular feedback behavior, which consumption trajectory segment the recommended product corresponds to as the support source, and which adjacent event reception relationships are included in that support source. After alignment, adjust the cross-channel intent reception coefficient. During adjustment, if a recommended product is clicked, added to cart, ordered, or redeemed with high trajectory support, increase the credibility level of its corresponding reception link. If a recommended product is ignored for a long time with high trajectory support, or a refund is generated after recommendation, decrease the credibility level of its corresponding reception link. Credibility level adjustment... The process is based on a preset feedback correction rule table, derived from the recommendation feedback management configuration, which specifies the correction magnitude of positive, negative, and neutral feedback on the connection relationship. Then, the member's dynamic preferences are updated. During the update, the direction of the feedback product's preference influence in terms of product category, brand, specifications, price, and promotional level is first identified. If the feedback behavior is a strongly positive behavior such as placing an order or redeeming a voucher, the preference item associated with the feedback product is strengthened. If the feedback behavior is a negative behavior such as ignoring or requesting a refund, the preference item associated with the feedback product is weakened. If the feedback behavior is merely a click or adding to cart, a moderate adjustment is made. Simultaneously, the stability of the consumption trajectory segment corresponding to the feedback product is checked; only when the trajectory support is high and the feedback direction is clear is a significant update of the member's dynamic preferences allowed. After the connection coefficient correction and preference update, the updated cross-channel intent connection coefficient and the updated member's dynamic preferences are obtained. The effect of this step is that the actual feedback after the recommendation output can flow back to the connection judgment and preference modeling, forming a continuous correction mechanism, thereby ensuring that subsequent recommendation results can be adjusted synchronously with changes in member behavior and connection relationships.

[0122] The present invention is further configured such that updating the cross-channel intent reception coefficient and member dynamic preferences based on the feedback records corresponding to the product recommendation results includes:

[0123] The system collects feedback records corresponding to product recommendation results, extracting exposure information, click information, add-to-cart information, order information, verification information, refund information, and ignore information to generate a feedback event sequence. Specifically, it reads the output product recommendation results and generates a unique recommendation identifier, recommendation time identifier, and recommendation location identifier for each recommended product. The recommendation identifier originates from the output records during the recommendation result generation stage and is used to map subsequent feedback to specific recommended products. The recommendation time identifier is used to define the starting point for feedback observation. The recommendation location identifier is used to distinguish feedback differences for the same product in different display locations. Subsequently, feedback records are collected around each recommendation identifier. Exposure information comes from the exposure log generated when the recommendation result is displayed. Click information comes from the interaction log generated when the user views or enters the details of the recommended product. Add-to-cart information comes from the operation log generated when the recommended product is added to the transaction intention set. Order information comes from the order log generated when the recommended product enters the valid order chain. Verification information comes from the completed fulfillment confirmation. The data includes performance logs generated when rights are redeemed; refund information comes from reverse transaction logs generated when completed goods are cancelled, returned, or offset; ignored information comes from silent results where recommended goods have been exposed but no clicks, add-to-cart, order placement, or redemption records have been generated within the preset observation window; the preset observation window is derived from the feedback collection rule table, set according to the effective observation period after recommendation display, used to distinguish between short-term delayed feedback and long-term no feedback; after collection, all types of feedback records are uniformly sorted according to the recommendation time, and each feedback record is supplemented with feedback type, feedback occurrence time, feedback intensity level, and feedback source link to generate a feedback event sequence; the feedback intensity level comes from the preset feedback classification table, where redemption and order placement are strong positive feedback, clicks and add-to-cart are intermediate feedback, refunds are strong negative feedback, and ignored is weak negative feedback; the effect of this step is to unify the feedback signals scattered in different business links into a traceable and comparable event sequence, providing continuous input for subsequent correlation and correction;

[0124] The feedback event sequence is correlated with product recommendation results, consumption trajectory fragments, and cross-channel intent carrying coefficients to generate feedback correlation results. Specifically, the feedback event sequence is first aligned one-to-one with the product recommendation results. During alignment, the recommendation identifier is used as the primary correlation key, and the recommendation time identifier and recommendation location identifier are used as auxiliary verification conditions to locate each feedback event to the corresponding recommended product. After completing the recommendation-level alignment, the feedback events are then correlated with consumption trajectory fragments. During correlation, the consumption trajectory support results corresponding to the recommended product in the recommendation generation stage are read to check which consumption trajectory fragment it comes from, which historical products are involved in the fragment, which unified event types, and which carrying links. Subsequently, the feedback events are linked with the cross-channel intent carrying coefficients. When establishing the link, along the consumption trajectory support path corresponding to the recommended product, adjacent consumption event pairs associated with the product are found, and then these adjacent consumption event pairs are read from the previous steps. The cross-channel intent reception coefficient is generated; then feedback attribution is performed, which refers to determining which reception link and which segment of preference deposition result a certain feedback event mainly affects; during attribution, the supporting link with the most direct product correspondence, the closest position in the consumption trajectory segment, and the highest reception stability is selected as the main attribution path; if multiple candidate supporting paths exist at the same time, they are filtered step by step according to the degree of product consistency, segment stability, and recommendation time proximity; after recommendation result alignment, trajectory segment association, reception link mapping, and feedback attribution, feedback association results are formed; the meaning of feedback association results is the result after unified organization of the recommended products, consumption trajectory segments, reception links, and direction of influence corresponding to each feedback event; the effect of this step is to clarify the feedback target and feedback path, so that subsequent reception coefficient correction and member dynamic preference updates have clear data sources and update boundaries;

[0125] Based on the feedback correlation results, the cross-channel intent acceptance coefficient is corrected to generate an updated cross-channel intent acceptance coefficient. Specifically, each acceptance link involved in the feedback correlation results is corrected separately. Before correction, the original cross-channel intent acceptance coefficient of the acceptance link is read, and the feedback event type, feedback intensity level, feedback time position, and segment stability level associated with the link are read. The segment stability level comes from the segment acceptance stability information in the previous consumption trajectory segment construction step and is used to reflect the overall acceptance quality of the segment in which the link is located. During correction, the preset feedback correction rule table is used. The feedback correction rule table comes from the feedback update configuration, which clearly stipulates that: if the feedback event is a redemption or order placement and the occurrence time falls within the effective observation window, the corresponding acceptance link is positively improved; if the feedback event is a click or add to cart, the corresponding acceptance link is moderately improved; if the feedback event is a refund, the corresponding acceptance link is significantly reduced; if the feedback event is ignored, the corresponding acceptance link is slightly reduced only after the recommended product has received sufficient exposure and exceeded the observation window. The correction magnitude is adjusted in conjunction with the segment stability level. If the segment of the associated link has a high stability level, the feedback result has a more direct impact on the acceptance coefficient. If the segment stability level is low, the correction magnitude of the feedback result is limited to avoid low-quality segments causing excessive disturbance to the acceptance relationship. For cases where multiple feedback events occur in the same acceptance link within an update cycle, they are first sorted by the time of occurrence, and then summarized and processed in a way that prioritizes strong feedback and allows subsequent feedback to cover preceding weak feedback. For example, if a click occurs before an order is placed, the subsequent order result takes precedence; if an order occurs before a refund, a reverse correction is performed according to the refund direction. After completing the above processing, the updated cross-channel intent acceptance coefficient is obtained. The meaning of the updated cross-channel intent acceptance coefficient is that, based on retaining the original acceptance judgment, a new value of acceptance strength is formed after absorbing the real feedback result. The effect of this step is that the acceptance coefficient can continuously reflect the verification result of the real feedback on the consumer intent link, thereby improving the self-correction capability of subsequent acceptance judgments.

[0126] Based on the updated cross-channel intent reception coefficient and feedback correlation results, the member dynamic preferences are updated to generate updated member dynamic preferences. Specifically, the current member dynamic preferences, the updated cross-channel intent reception coefficient, and the feedback correlation results are read. Member dynamic preferences are derived from prior preference deposition and long-term and recent fusion results, and internally include at least product category preferences, brand preferences, specification preferences, price level preferences, and promotion response preferences. During the update, the direction of feedback effect is first determined. If the feedback event is a redemption or order placement, the feedback effect is positive reinforcement; if the feedback event is a click or add to cart, the feedback effect is moderate reinforcement; if the feedback event is a refund, the feedback effect is negative weakening; if the feedback event is ignored, the feedback effect is slight weakening. Then, the scope of feedback influence is determined; the scope of feedback influence is derived from the product correspondence and trajectory support path recorded in the feedback correlation results. If the recommended product and the historically consumed product belong to the same unified product index, the preference item to which the product belongs will be directly affected. If there is a specification relationship or packaging relationship between the recommended product and the historically consumed product, the corresponding specification preference item or packaging preference item will be affected simultaneously. If the recommended product comes from a combination relationship or fulfillment connection relationship extension, the combination preference item or fulfillment preference item will be affected simultaneously. Then, the update weight is determined, which is determined by three factors: First, the feedback intensity level, strong feedback corresponds to a higher update magnitude, medium feedback corresponds to a medium update magnitude, and weak feedback corresponds to a lower update magnitude; second, the cross-channel intent inheritance coefficient after the update, the higher the inheritance coefficient, the stronger the connection between the recommendation result and the existing consumption intent chain, and the higher the corresponding update weight; third, the recentity of the feedback time, the closer the feedback is to the current update time. The greater the impact of a point on recent preferences, the more significant its influence. After determining the updated weights, adjustments are made to the corresponding preference items in the member's dynamic preferences, while simultaneously checking the consistency between long-term and recent preferences. If current feedback primarily affects recent consumption, recent preferences are corrected first. If similar feedback recurs in multiple update cycles, it is gradually transmitted to long-term preferences. Finally, all updated preference items are consolidated to maintain the stability of the hierarchical order and weight distribution within the preference structure, generating updated member dynamic preferences. The effect of this step is that member dynamic preferences can be continuously corrected under the drive of real feedback, and the update magnitude is controlled by three factors: the acceptance coefficient, feedback intensity, and time position, thereby ensuring that the source of preference changes is clear, the update path is clear, and the correction results are traceable.

[0127] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for member preference modeling and product recommendation based on omnichannel consumption patterns, characterized in that, include: Collect members' channel consumption records, perform time correction, product mapping and event normalization on the channel consumption records, and generate a unified consumption event sequence carrying identity clues and event context information; A unified member index is generated by performing association parsing based on identity clues in the unified consumption event sequence; Based on the unified member index, the unified consumption event sequence is rearranged, and the event context information is judged to determine the continuity. The cross-channel intent continuity coefficient and consumption trajectory fragment are output. Execute preference deposition on consumption trajectory segments, and generate dynamic member preferences by combining cross-channel intent acceptance coefficients; Collect product operation records, and organize and associate the execution status of product operation records to generate product constraint information; Based on consumption trajectory fragments and product constraint information, correlation filtering is performed, and product recommendation results are generated by combining dynamic member preferences. 2.The method of claim 1, wherein, Generating a unified sequence of consumption events carrying identity clues and event context information includes: Channel consumption records include cashier records, online transaction records, platform interface records, and fulfillment records; The channel consumption records are decomposed into fields, and the original time information, channel source information, store information, product information, action information, order association information, payment association information, promotion information, fulfillment association information and identity information are extracted and encapsulated according to the preset field structure to generate the original event carrier; Based on order association information, payment association information, and performance association information, a unified timestamp is generated by aligning and offsetting the execution time of the original event carrier. Perform matching constraints and hierarchical validations on product information against a pre-defined standard product database to generate a unified product index; By combining action information, order association information, payment association information, and fulfillment association information, the execution events of the original event carrier are normalized to generate a unified event type. The channel source information, store information, promotion information, fulfillment association information, and unified timestamp are organized to generate event context information. Based on the identity information, a reliability screening process is performed to generate identity clues; A unified consumption event sequence is generated by encapsulating and sorting unified timestamps, unified product indexes, unified event types, event context information, and identity clues.

3. The method for member preference modeling and product recommendation based on omnichannel consumption trajectory as described in claim 1, characterized in that, Based on the identity clues in the unified consumption event sequence, a unified member index is generated by performing association resolution, including: Based on the identity clues in the unified consumption event sequence, type sorting, format standardization, and semantic regularization are performed to generate identity clue groups divided by telephone number identifier, payment identifier, device identifier, account identifier, and address identifier; By combining the event context information in the unified consumption event sequence, source reliability verification, format integrity verification, and context consistency verification are performed on the identity clue group to generate the event-level identity description of the corresponding unified consumption event; Based on identity clue groups and event-level identity descriptions, candidate association filtering and association scoring are performed on consumption events in a unified consumption event sequence to generate a candidate association set; Merge and aggregate the candidate association set and perform conflict checks to generate member merge groups. Then, perform index allocation on the member merge groups to generate a unified member index.

4. The method for member preference modeling and product recommendation based on omnichannel consumption trajectory according to claim 3, characterized in that, The unified consumption event sequence is rearranged based on the unified member index, and succession judgment is performed on the event context information, including: Based on the member merging groups corresponding to the unified member index, the unified consumption event sequence is divided by affiliation and sorted by time sequence to generate member event subsequences; Perform candidate construction and span filtering on adjacent consumption events in the member event subsequence to generate valid successor candidate event pairs; Extract the event context information of the preceding and subsequent consumption events in the valid candidate event pair, perform time interval calculation, channel migration identification, promotion link comparison and fulfillment status connection determination, and generate a context description. Extract the unified product index, unified event type, and store information of the preceding and subsequent consumption events in the valid candidate event pairs, perform product correspondence comparison, event flow order determination, and store acceptance determination, and generate a continuous semantic description of the events. 5.The method of claim 4, wherein, Output cross-channel intent reception coefficients and consumption trajectory fragments, including: A joint determination is performed on the context description and the continuous semantic description of the event to generate the cross-channel intent acceptance coefficient for the effective acceptance candidate event pair; Based on the cross-channel intent acceptance coefficient, the execution order of adjacent consumption events in the member event subsequence is segmented to generate initial consumption trajectory fragments; Perform internal continuity checks and boundary adjustments on the initial consumption trajectory segment to generate a new consumption trajectory segment. 6.The method of claim 1, wherein, The system performs preference deposition on consumption trajectory segments and generates dynamic member preferences by combining cross-channel intent reception coefficients, including: Based on consumption events in consumption trajectory segments, a unified product index, a unified event type, and event context information are extracted, and preference semantic mapping is performed to generate event preference primitives. By combining cross-channel intent acceptance coefficients, unified event types, and time information, the event deposition intensity is used to determine the generated deposition weights. The event preference primitives and deposition weights are sequentially deposited to generate fragment preference representations, and the cross-channel intent acceptance coefficients within the consumption trajectory fragments are organized to generate fragment acceptance stability information. Based on fragment preference representation, fragment-inherited stable information and temporal information, cross-fragment fusion is performed to generate long-term and short-term preferences; Dynamic preferences for members are generated by dynamically fusing long-term and short-term preferences. 7.The method of claim 1, wherein, Collect product operation records, and generate product constraint information by normalizing and associating the execution status of these records, including: Collect product operation records and extract product identifiers, specification information, channel operation identifiers, store operation identifiers, operation time, inventory information, price information, promotion execution information, fulfillment execution information, and source information to generate product operation atoms; Product identification and specification information are mapped to a unified product index space. Time window slicing is performed on the operation time. Based on the unified product index, channel operation identification, store operation identification and slice time, the atomic execution of product operation is aggregated to generate scenario product units. The execution status of inventory information, price information, promotion execution information and fulfillment execution information in the scene product unit is unified, and the execution status quantity is generated by merging and integrating the source information. Based on the specification relationships, packaging relationships, combination relationships and fulfillment connection relationships corresponding to the unified product index, the scene product units are associated and organized to generate associated description quantities. The unified product index, channel operation identifier, store operation identifier, slice time, scene state quantity, and associated description quantity are encapsulated to generate product constraint information. 8.The method of claim 7, wherein, The association filtering based on consumption trajectory fragments and product constraint information includes: Extract the unified product index, unified event type, and event context information from the consumption trajectory segment, and match them with the unified product index, channel operation identifier, store operation identifier, and slice time in the product constraint information to generate candidate product items; Based on the inventory status, price status, promotion status, and fulfillment status in the scenario state quantity, constraint filtering is performed on candidate product items to generate an initial candidate product set; Based on the specification relationships, packaging relationships, combination relationships, and fulfillment connection relationships in the associated descriptive quantities, the initial candidate product set is expanded and deduplicated to generate a new candidate product set. 9.The method of claim 8, wherein, Product recommendations generated based on members' dynamic preferences include: Perform preference matching between members' dynamic preferences and the candidate product set to generate preference matching results corresponding to the candidate products; By combining consumption trajectory fragments and cross-channel intent reception coefficients, the execution trajectory support for the candidate product set is determined, and the trajectory support results corresponding to the candidate products are generated. Based on preference matching results and trajectory support results, the candidate product set is recommended and ranked to generate product recommendation results; Based on the feedback records corresponding to the product recommendation results, update the cross-channel intent reception coefficient and member dynamic preferences. 10.The method of claim 9, wherein, Based on the feedback records corresponding to the product recommendation results, update the cross-channel intent reception coefficient and member dynamic preferences, including: Collect feedback records corresponding to product recommendation results, extract exposure information, click information, add-to-cart information, order information, redemption information, refund information, and ignore information, and generate a feedback event sequence; The feedback event sequence is correlated with product recommendation results, consumption trajectory fragments, and cross-channel intent reception coefficients to generate feedback correlation results. Based on the feedback correlation results, the cross-channel intent acceptance coefficient is corrected to generate an updated cross-channel intent acceptance coefficient; Based on the updated cross-channel intent reception coefficient and feedback correlation results, the member dynamic preferences are updated to generate updated member dynamic preferences.

Images