Object recall method, system and apparatus

Abstract

Embodiments of the present specification provide an object recall method, system and device, wherein the method comprises: in response to a recall request, determining a plurality of recall paths corresponding to the recall request and path priority information of the plurality of recall paths, wherein the plurality of recall paths at least includes a behavior triggered recall path; according to the path priority information, sequentially performing object recall according to the plurality of recall paths to obtain a recall object set corresponding to the recall request; wherein performing object recall according to the behavior triggered recall path comprises: determining a user behavior sequence corresponding to the recall request, and determining a plurality of recall levels corresponding to the behavior triggered recall path and a level priority order of the plurality of recall levels according to the user behavior sequence; according to the level priority order, sequentially performing object recall according to the plurality of recall levels to obtain a recall object corresponding to the behavior triggered recall path and add the recall object to the recall object set. Through a double-layer hierarchical recall mechanism, the relevance and quality of the recall result are improved.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of search and retrieval technology, and in particular to an object retrieval method, system and apparatus. Background Technology

[0002] With the rapid development of e-commerce, the number of products on e-commerce platforms has grown exponentially, making search systems a key entry point connecting users and products. Search systems typically consist of two phases: recall and ranking. The recall phase is responsible for quickly filtering a set of candidate products from the entire product database for processing in the subsequent ranking phase.

[0003] Currently, multi-path recall strategies are widely used in industrial-grade search systems. Multi-path recall employs multiple recall strategies with different logics simultaneously to mine products relevant to user queries from different perspectives. The results from each path are then merged and sent to the ranking stage. However, in practical applications, the recall effectiveness of different strategies varies, resulting in inconsistent quality of the recall results. Existing systems typically merge the results from each path and process them uniformly, lacking an effective differentiated quality control mechanism. This leads to the need to process a large number of low-quality candidates during the ranking stage, increasing the system's computational overhead.

[0004] Therefore, how to effectively organize and manage multi-pronged recall strategies and improve the overall effectiveness of the recall phase is an urgent problem to be solved. Summary of the Invention

[0005] In view of this, embodiments of this specification provide an object recall method. One or more embodiments of this specification also relate to an object recall system, an object recall device, a computing device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0006] According to a first aspect of the embodiments of this specification, an object recall method is provided, comprising: In response to a recall request, multiple recall paths corresponding to the recall request and path priority information of the multiple recall paths are determined, wherein the multiple recall paths include at least a behavior-triggered recall path; Based on the path priority information, objects are recalled sequentially according to multiple recall paths to obtain the set of recalled objects corresponding to the recall request; The process of recalling objects according to the behavior-triggered recall path includes: determining the user behavior sequence corresponding to the recall request, and determining multiple recall levels corresponding to the behavior-triggered recall path based on the user behavior sequence, as well as the priority order of the multiple recall levels; and sequentially recalling objects according to the priority order of the multiple recall levels to obtain the recalled objects corresponding to the behavior-triggered recall path and add them to the recall object set.

[0007] According to a second aspect of the embodiments of this specification, an object recall system is provided, the system comprising a client and a server, wherein... The client, in response to the search command issued by the user for the search page, generates a recall request and sends it to the server; The server executes object recall according to the above object recall method, obtains the recall object set corresponding to the recall request, and returns it to the client. The client displays the recalled objects corresponding to the recalled object set on the search page.

[0008] According to a third aspect of the embodiments of this specification, an object recall apparatus is provided, comprising: The determination module is configured to, in response to a recall request, determine multiple recall paths corresponding to the recall request and path priority information of the multiple recall paths, wherein the multiple recall paths include at least a behavior-triggered recall path. The acquisition module is configured to sequentially retrieve objects according to multiple retrieval paths based on the path priority information, and obtain the set of recalled objects corresponding to the retrieval request. The process of recalling objects according to the behavior-triggered recall path includes: determining the user behavior sequence corresponding to the recall request, and determining multiple recall levels corresponding to the behavior-triggered recall path based on the user behavior sequence, as well as the priority order of the multiple recall levels; and sequentially recalling objects according to the priority order of the multiple recall levels to obtain the recalled objects corresponding to the behavior-triggered recall path and add them to the recall object set.

[0009] According to a fourth aspect of the embodiments of this specification, a computing device is provided, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the object recall method described above.

[0010] According to a fifth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions that, when executed by a processor, implement the steps of the object recall method described above.

[0011] According to a sixth aspect of the embodiments of this specification, a computer program product is provided, including a computer program or instructions that, when executed by a processor, implement the steps of the object recall method described above.

[0012] One embodiment of this specification implements object recall by sequentially following multiple recall paths based on path priority information. This achieves differentiated processing for different recall paths, ensuring that recall results from high-priority recall paths are prioritized for inclusion in the recall object set. Low-priority recall paths are only used as supplementary paths when the number of recalled objects is insufficient, preventing low-quality recall results from crowding out the recall space. Simultaneously, within the behavior-triggered recall path, multiple recall levels are determined based on the user behavior sequence, and object recall is performed sequentially according to level priority. This achieves fine-grained control within the same recall path, prioritizing recall levels with higher relevance to the user behavior sequence and avoiding the problem of insufficient relevance caused by single-level recall. The aforementioned hierarchical recall path and intra-path recall layering work synergistically to construct a complete quality control system from both inter-path and intra-path dimensions. While ensuring priority exposure of highly relevant recall results, it rationally utilizes low-relevance levels to supplement recall coverage, thereby improving the overall effectiveness of the recall phase and the relevance of recall results to recall requests. This reduces the number of low-quality candidate objects entering subsequent sorting stages and lowers ineffective computational overhead. Attached Figure Description

[0013] Figure 1 A schematic diagram of the architecture of an object recall method according to an embodiment of this specification is shown; Figure 2 A flowchart of an object recall method according to an embodiment of this specification is shown; Figure 3 A flowchart illustrating the processing procedure of an object recall method provided in one embodiment of this specification is shown. Figure 4 This specification shows a schematic diagram of the structure of an object recall system provided in one embodiment; Figure 5 This specification shows a schematic diagram of the structure of an object recall device according to one embodiment; Figure 6 A structural block diagram of a computing device provided according to one embodiment of this specification is shown. Detailed Implementation

[0014] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0015] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0016] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0017] Furthermore, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0018] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0019] Entity recall: Entity recall refers to the process of filtering candidate objects from an object database that match a recall request. In the embodiments of this specification, entity recall can be used to quickly retrieve products that may be related to the user's query from the full product database, as a candidate set for the subsequent ranking stage.

[0020] Recall Path: A recall path refers to a technical channel that uses specific matching logic to recall objects. In the embodiments of this specification, different recall paths employ different recall strategies, mining objects related to the recall request from different dimensions. The results of each recall path are ultimately merged to form a set of recalled objects.

[0021] I2I Recall: I2I recall (Item-to-Item) refers to a recall method that starts with a user's historical behavior object and uses a two-hop association between objects to obtain a recall object associated with the triggering object. Here, "first hop" refers to the interaction between the user and the historical behavior object, and "second hop" refers to finding similar objects along the associations between the historical behavior object. In the embodiments of this specification, I2I recall is a specific implementation of the behavior-triggered recall path.

[0022] S2I Recall: S2I recall (Store-to-Item) refers to the method of recalling products from a store based on the user's preferred store information. In the embodiments of this specification, S2I recall is used as one of the multiple recall methods to recall objects based on the user's store preferences.

[0023] B2I Recall: B2I recall (Brand-to-Item) refers to the method of recalling products from a brand based on the user's preferred brand information. In the embodiments of this specification, B2I recall is used as one of the multiple recall methods to recall objects based on the user's brand preferences.

[0024] Q2I Recall: Q2I recall (Query-to-Item) refers to the path of retrieving items by matching them with the text information contained in the recall request. In the embodiments of this specification, Q2I recall is based on the relevance between the query terms and the object text.

[0025] C2I Recall: C2I recall (Category-to-Item) refers to the path of recalling objects based on the target object category corresponding to the recall request. In the embodiments of this specification, C2I recall is based on category matching for object recall.

[0026] User Trigger: A user trigger (trigger object) refers to a behavior object in a user's historical behavior sequence that serves as the starting point for recall. In the embodiments of this specification, the user trigger is used as a starting node in the behavior trigger recall path to conduct diffusion recall through the association relationship between objects.

[0027] Two-hop attachment: Two-hop attachment refers to the process of finding related objects by starting from the triggering object and following the association relationships in an object relationship network. In the embodiments of this specification, two-hop attachment is used to implement object diffusion recall in the behavior trigger recall path.

[0028] With the rapid development of e-commerce, the number of products on e-commerce platforms has grown exponentially. Search systems have become the core entry point connecting users and products, with the core objective of quickly and accurately filtering products from a massive product database that match the user's query needs. Search systems typically consist of two phases: recall and ranking. The recall phase is responsible for quickly filtering a set of candidate products relevant to the user's query from the entire product database, which is then used by the ranking phase for refined scoring and sorting. The effectiveness of the recall phase directly impacts the input quality of the ranking phase, thus affecting the final performance of the entire search system.

[0029] In the recall phase, multi-path recall strategies are widely used in industrial-grade search systems. Multi-path recall refers to the system simultaneously employing multiple recall strategies with different logics to mine objects related to user queries from different perspectives, thereby improving recall coverage and relevance. Common recall paths include text-based recall paths, which match user-input text with object text; category-based recall paths, which filter objects based on the category corresponding to the user's query; user behavior-based recall paths, which expand recall based on the relationships between objects based on the user's historical behavior; store-preference-based recall paths, which filter objects based on the user's preferred stores; and brand-preference-based recall paths, which filter objects based on the user's preferred brands. These recall paths work together from multiple dimensions such as keywords, categories, behaviors, stores, and brands. The results of all recall paths are ultimately merged and sent to the ranking phase.

[0030] However, in practical applications, the recall effects of different recall paths vary objectively. Due to the different logical mechanisms of each recall path, the quality, relevance, and conversion contribution of their recall results often differ. Furthermore, within a single path, especially in recall paths based on user behavior association, the diversity of users' historical behaviors can easily lead to recalls triggered by behaviors with low relevance to the current query intent, resulting in a discrepancy between the recall results and the user's query intent. Existing systems typically merge the recall results from all paths and send them uniformly to the ranking stage, lacking an effective differentiated quality control mechanism. This equal-weighting approach results in the ranking stage needing to process a large number of low-quality candidates, not only increasing the system's inefficient computational overhead but also potentially leading to insufficient exposure of candidates with higher relevance to the query intent. Therefore, how to effectively organize and manage multi-path recall strategies and improve the overall effectiveness and relevance of the recall stage is a pressing issue that needs to be addressed.

[0031] Based on this, this specification provides an object recall method, and also relates to an object recall system, an object recall device, a computing device, a computer-readable storage medium, and a computer program product, which will be described in detail in the following embodiments.

[0032] See Figure 1 , Figure 1 A schematic diagram of the architecture of an object recall method according to one embodiment of this specification is shown. Figure 1 As shown, the object recall method provided in the embodiments of this specification adopts a two-layer hierarchical architecture, including recall path layering and recall level layering.

[0033] In the hierarchical recall path system, in response to a recall request, multiple recall paths corresponding to the request and their priority information are determined. These multiple recall paths include at least a behavior-triggered recall path. Objects are recalled sequentially through these paths according to their priority information, with higher-priority paths executed first and lower-priority paths executed as supplementary steps when necessary. This hierarchical recall path system enables differentiated processing of each recall path, avoiding low-quality recall results that crowd out the recall space due to competition among all recall paths.

[0034] In the hierarchical recall mechanism, for each behavior-triggered recall path, the user behavior sequence corresponding to the recall request is determined. Based on this sequence, multiple recall levels corresponding to the path and their priority order are established. Objects are then recalled sequentially through these levels according to their priority. Higher-priority levels are executed first, while lower-priority levels are executed only when higher-level conditions are not met. This hierarchical approach allows for fine-grained control over the behavior-triggered recall path, preventing semantic drift caused by behavior-triggered objects that do not match the query intent. The hierarchical recall path and level structures work together to form a complete object recall architecture. After execution, the objects recalled by each path and level are added to the recall object set as input for subsequent sorting stages.

[0035] By layering the recall path, differentiated processing of different recall paths is achieved, preventing low-quality recall results from crowding out the recall space. By implementing hierarchical recall levels within the behavior-triggered recall path, fine-grained control within the same path is achieved, improving the relevance of recall results to query requests. The synergistic cooperation between recall path layering and recall level layering ensures priority exposure of highly relevant recall results while rationally utilizing lower-relevance levels to supplement recall coverage. This reduces the number of low-quality candidates entering subsequent ranking stages, lowers ineffective computational overhead, and improves the overall effectiveness of the recall phase.

[0036] See Figure 2 , Figure 2 A flowchart of an object recall method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0037] Step 202: In response to the recall request, determine multiple recall paths corresponding to the recall request and path priority information of the multiple recall paths, wherein the multiple recall paths include at least a behavior-triggered recall path.

[0038] In this context, a recall request refers to a user-initiated object search request, which may include user-inputted text information such as search terms and user identification information. The recall request triggers the entire object recall process. A recall path refers to the technical channel that uses specific matching logic for object recall. Different recall paths employ different recall strategies, such as text-based recall paths (Q2I), category-based recall paths (C2I), user behavior-triggered recall paths (I2I), store-preference-based recall paths (S2I), and brand-preference-based recall paths (B2I). Each recall path mines objects related to the recall request from different dimensions. Path priority information refers to the execution priority order among the recall paths. In the embodiments of this specification, path priority information is pre-determined based on the historical hit data of each recall path and is used to control the execution order of multiple recalls. For example, recall paths are divided into high-priority, medium-priority, and low-priority tiers, and are executed sequentially from highest to lowest priority.

[0039] A behavior-triggered recall path refers to a path that uses a user's historical behavior sequence as the starting point and recalls objects through pre-defined relationships between objects. This path differs from recall paths based on text matching or category matching. Its core lies in utilizing the user's historical behavior data, such as clicks, add-to-cart actions, and purchases, to uncover the user's potential interests, thereby achieving personalized recall. In the embodiments of this specification, one specific implementation of the behavior-triggered recall path is an I2I recall path, which starts from a user's historical behavior object and uses pre-defined relationships between objects to perform a two-hop attachment to obtain recall objects associated with the trigger object. It should be noted that the behavior-triggered recall path in this specification is not limited to the I2I path; other recall methods that use the user's historical behavior as the starting point are also applicable.

[0040] In practical applications, during search and retrieval scenarios, each user-initiated retrieval request requires determining which retrieval paths to use and the execution order of each path. However, the retrieval quality and effectiveness of different paths vary objectively. Existing systems treat all retrieval paths with equal weight, failing to consider efficiency differences between paths. This results in low-quality retrieval paths competing equally with high-quality paths during the retrieval phase, increasing unnecessary computational overhead and potentially leading to insufficient exposure of high-quality candidates. Therefore, in this embodiment, in response to a retrieval request, multiple retrieval paths corresponding to the request and the path priority information of each path are determined. This provides a basis for subsequently processing different retrieval paths differently according to priority, ensuring that the organization of multi-path retrieval is targeted and differentiated.

[0041] In practice, recall requests may also originate from non-text search scenarios such as users clicking on recommended sections or viewing similar recommendations on product detail pages. In these scenarios, the recall request may not contain user-inputted text information, but rather information such as the identifier of the object currently being viewed or the identifier of the current category. Accordingly, when determining multiple recall paths, the system can determine the applicable recall path based on the object identifier or category identifier carried in the request. For example, in a similar recommendation scenario, the user may not input text information, and the system can directly trigger a C2I recall path based on the category of the currently viewed object and an I2I recall path based on the currently viewed object. The recall requests described in one or more embodiments of this specification are not limited to text search requests, but can also be any other scenario request that triggers object recall.

[0042] Furthermore, path priority information can be pre-determined through offline evaluation. Specifically, the system can perform offline efficiency evaluations on multiple candidate recall paths, such as using metrics like click-through rate or conversion rate to quantify the recall quality of each path. Based on the evaluation results, the system ranks the candidate recall paths and divides them into at least two priority tiers: high priority, medium priority, and low priority. Recall paths with higher evaluation results correspond to the high priority tier, and those with lower results correspond to the low priority tier. The mapping relationship between each recall path and priority tier is configured as information that can be read online, allowing the system to directly determine the path priority information of each recall path based on the pre-configured mapping relationship.

[0043] When determining multiple recall paths, as an optional approach, the system can obtain the request information of the recall request and / or the user's historical behavior information corresponding to the recall request, and determine multiple recall paths corresponding to the current recall request based on the request information and / or the user's historical behavior information. For example, if the recall request contains the user's input search term, a text-based recall path is triggered; if the category prediction of the search term can obtain the target object category, a category-based recall path is triggered; if the user has a historical behavior sequence, a behavior-based recall path is triggered; if the user has store preference information, a store preference-based recall path is triggered, and so on. In this way, different recall requests can dynamically adapt to different recall path combinations according to their own characteristics, realizing on-demand configuration of recall paths.

[0044] In a specific embodiment of this specification, a user enters "red dress" in the search box of an e-commerce app. After receiving the recall request, the system first obtains the user's search term information and user identifier information. Based on the search term "red dress," the system predicts that the target object category is "dress," and determines the available recall paths, including: text-matching-based recall paths, category-matching-based recall paths, user behavior-triggered recall paths, and store preference-based recall paths. Simultaneously, the system looks up the priority level corresponding to each recall path from a pre-configured path priority information table. User behavior-triggered recall paths correspond to high priority levels, category-matching-based recall paths correspond to medium priority levels, and text-matching-based and store preference-based recall paths correspond to low priority levels. This determines the multiple recall paths corresponding to this recall request and the path priority information of each path.

[0045] Based on this, by responding to recall requests and determining the multiple recall paths corresponding to each request and the path priority information of each path, the applicable recall path can be dynamically adapted according to the specific circumstances of each recall request. This avoids using a fixed combination of recall paths for all requests, thereby improving the matching degree between the recall strategy and the request scenario. At the same time, by introducing path priority information, a differentiated execution order basis is established for each recall path, avoiding low-quality recall results caused by competition among all paths with equal weight, which crowds out the recall space. This lays the foundation for the subsequent execution of each recall path in order of priority, which helps to improve the overall efficiency and quality of the recall phase.

[0046] Furthermore, in response to the recall request, the method further includes: determining multiple candidate recall paths and historical hit data for each candidate recall path; sorting the multiple candidate recall paths according to the historical hit data for each candidate recall path to obtain multiple sorted candidate recall paths; selecting multiple recall paths from the multiple sorted candidate recall paths according to a preset path priority number, and setting path priority information corresponding to each recall path.

[0047] Candidate recall paths refer to the set of recall paths available for the system to choose from. These paths include, but are not limited to, text-matching-based, category-matching-based, user behavior-triggered, store-preference-based, and brand-preference-based recall paths. Candidate recall paths form the basis for the system's path selection and priority configuration.

[0048] Historical hit data refers to the hit rate metric of a recall path in offline evaluation, used to quantify the recall quality or efficiency of each recall path. Historical hit data can include click-through rate (CTR) and / or conversion rate (CR). Specifically, CTR refers to the proportion of users whose actual clicks were among the candidate objects recalled by a recall path in offline evaluation; CR refers to the proportion of users whose actual purchases were among the candidate objects recalled by a recall path in offline evaluation. Higher historical hit data indicates higher recall quality or efficiency for that recall path.

[0049] The number of preset path priorities refers to the number of priority levels preset by the system. In the embodiments of this specification, the number of preset path priorities can be three, namely high priority level, medium priority level and low priority level, but it is not limited to this, and can also be two or more levels.

[0050] In practical applications, the system needs to determine the path priority information of each recall path in response to a recall request, but this requires pre-establishing a mapping relationship between each recall path and its priority level. However, how to objectively and quantitatively measure the recall quality differences of each recall path, and how to reasonably determine the priority order of each path based on the measurement results, are prerequisite problems that need to be solved. If the division is based solely on manual experience, it will be subjective and uncertain, making it difficult to accurately reflect the actual recall effect differences of each path. Therefore, before responding to a recall request, a quantitative ranking is performed based on the historical hit data of each candidate recall path, and the path priority information of each recall path is set according to the ranking results, thus providing an objective and quantitative basis for determining the path priority.

[0051] In practice, the historical timeframe used for offline evaluation can be user behavior data from the past 7, 15, or 30 days, depending on project requirements and data volume. Within the historical hit data, click-through rate (CTR) focuses on the match between the retrieved results and user click interests, while conversion rate (CTR) focuses on the match between the retrieved results and user purchase conversion behavior. Both can be used individually or in combination. When used together, the weighting coefficients of CTR and CTR can be dynamically adjusted based on project goals, such as improving CTR or improving conversion rates. The specific values ​​mentioned above are for illustrative purposes only; in actual applications, they should be determined based on the project scenario and data distribution.

[0052] It should be noted that, as an optional method, when sorting based on the historical hit data of each candidate recall path, for multiple candidate recall paths with the same or similar evaluation results, such as hit rate differences within a preset threshold range, the system can group them into the same ranking position and assign them the same priority level accordingly. This can avoid unnecessary priority differences caused by minor data fluctuations and improve the robustness of the system.

[0053] In addition, as an alternative approach, among the multiple candidate recall paths after sorting, the system is not limited to selecting the same number of recall paths as the preset path priority level. Instead, it can select more or fewer recall paths based on actual project needs. For example, even if three priority levels are preset, the system can select only the top two recall paths as available online, or select the top four recall paths and group two of them into the same priority level. This manual does not impose any limitations on this; the specific configuration can be flexibly tailored to actual application requirements.

[0054] In a specific embodiment of this specification, the system selects user behavior data from a historical time period as a test set during the offline phase to obtain a full set of candidate recall paths, including recall paths triggered by user behavior, recall paths based on category matching, recall paths based on text matching, recall paths based on store preferences, and recall paths based on brand preferences. The system performs offline evaluation on each candidate recall path, calculates the click-through rate and conversion rate of each path, and obtains historical hit data for each path. Assuming the evaluation results are: the click-through rate of the recall path triggered by user behavior is 48.24%, the click-through rate of the recall path based on category matching is 31.04%, the click-through rate of the recall path based on text matching is 28.59%, the click-through rate of the recall path based on store preferences is 9.44%, and the click-through rate of the recall path based on brand preferences is 15.23%. The system sorts candidate recall paths from highest to lowest click-through rate. The sorting result is as follows: user behavior-triggered recall paths rank first, category matching-based recall paths rank second, text matching-based recall paths rank third, brand preference-based recall paths rank fourth, and store preference-based recall paths rank fifth. The system pre-sets three priority levels: high priority, medium priority, and low priority. Therefore, according to the sorting order, the user behavior-triggered recall path (ranked first) is set to high priority, the category matching-based recall path (ranked second) is set to medium priority, and the text matching-based recall path (ranked third) is set to low priority. This mapping relationship is configured as information that can be read from an online lookup table so that when responding to a recall request, the system can determine the path priority information of each recall path corresponding to the current request based on this mapping relationship.

[0055] Based on this, by sorting the historical hit data of each candidate recall path before responding to the recall request and setting the path priority information of each recall path according to the sorting results, the recall efficiency of each recall path can be objectively and quantitatively evaluated. This ensures that the priority setting is supported by data rather than subjective human experience, thereby guaranteeing the rationality and accuracy of the path priority division. At the same time, by pre-setting the number of path priorities, each sorted path is assigned to a different priority level, realizing hierarchical management of each recall path. This provides a pre-configuration basis for executing each recall path differently according to priority, thereby ensuring that high-efficiency recall paths have the opportunity to gain priority exposure in online recall.

[0056] Step 204: Based on the path priority information, retrieve objects sequentially according to multiple retrieval paths to obtain the set of retrieved objects corresponding to the retrieval request; The process of recalling objects according to the behavior-triggered recall path includes: determining the user behavior sequence corresponding to the recall request, and determining multiple recall levels corresponding to the behavior-triggered recall path based on the user behavior sequence, as well as the priority order of the multiple recall levels; and sequentially recalling objects according to the priority order of the multiple recall levels to obtain the recalled objects corresponding to the behavior-triggered recall path and add them to the recall object set.

[0057] Here, path priority information refers to the execution order among various recall paths. In this step, path priority information is used to control the execution order of multiple recall paths, ensuring that higher-priority recall paths are executed first, and lower-priority recall paths are executed later. User behavior sequence refers to a collection of historical user behavior data arranged in chronological order. This historical behavior data includes the behavior objects corresponding to user clicks, adding to cart, purchasing, and exposure, along with their timing information. In this step, user behavior sequence is used to provide the trigger object source for behavior-triggered recall paths.

[0058] Recall tiers refer to different recall levels within the behavior-triggered recall path, based on the relevance between the category of the triggering object and the category of the target object. Different recall tiers correspond to different levels of category relevance. In this step, recall tiers are used for fine-grained hierarchical control of recalls within the behavior-triggered recall path. Tier priority order refers to the execution order among different recall tiers. In this step, the tier priority order is determined according to the relevance of each tier to the target object category, from highest to lowest. The higher the relevance, the higher the priority of the recall tier, and the earlier it is executed. The recall object set refers to the final set of candidate objects obtained for processing in the subsequent sorting stage. In this step, objects recalled by each recall path and each recall tier are added to this recall object set.

[0059] In practical applications, after determining multiple recall paths and their priority information, the problem to be solved is how to execute each path in an orderly manner according to priority for object recall. If all recall paths are executed simultaneously and the results are simply merged, the low-quality recall results of low-priority paths may be mixed equally with the high-quality recall results of high-priority paths, failing to reflect the efficiency differences between different paths. Furthermore, for behavior-triggered recall paths, due to the diversity of user historical behaviors, some behaviors may have low semantic relevance to the current recall request. If all are indiscriminately used as trigger objects for recall, semantic drift in the recall results can easily occur. Therefore, in the embodiments of this specification, on the one hand, each recall path is executed sequentially according to the path priority information to ensure that high-priority paths are prioritized for recall; on the other hand, within the behavior-triggered recall path, multiple recall levels are determined based on the user behavior sequence and executed sequentially according to the level priority to ensure that behavior trigger objects with higher semantic relevance to the recall request are prioritized for recall, thereby achieving dual orderly control at the external path level and internal level.

[0060] In practical implementation, as an optional approach, when recalling objects sequentially according to multiple recall paths based on path priority information, the system can execute each recall path in descending order of priority and accumulate the number of recalled objects. When the accumulated number reaches the preset target recall total, the execution of the current priority and all lower priority recall paths is stopped. This avoids unnecessary computational overhead caused by performing recall on all paths.

[0061] Furthermore, within the behavior-triggered recall path, when determining multiple recall levels based on the user behavior sequence, as an optional approach, the system can first determine the target object category corresponding to the recall request, and the object category of each behavior object in the user behavior sequence; then, based on the target object category, the object categories of each behavior object are divided into multiple object category groups, where each group has a different degree of association with the target object category; and multiple recall levels and their priority order are determined based on each group and its degree of association. Specifically, behavior objects whose object category matches the target object category can be identified as trigger objects for the first recall level; behavior objects whose object category belongs to the same parent category as the target object category can be identified as trigger objects for the second recall level; and behavior objects whose object category does not belong to the same parent category as the target object category can be identified as trigger objects for the third recall level. The priority of the first recall level is higher than that of the second recall level, and the priority of the second recall level is higher than that of the third recall level.

[0062] Within the behavior-triggered recall path, when recalling objects sequentially according to hierarchical priority levels, each recall level can be executed in descending order of priority, and the number of objects recalled at each level can be accumulated. When the number of objects recalled at a certain recall level reaches the preset recall quantity for that level, execution of the next recall level is stopped; or, after all levels have been executed, the objects recalled at each level are merged as the recall result of the behavior-triggered recall path.

[0063] In a specific embodiment of this specification, a user searches for a red dress. The system has determined multiple recall paths and their priority information, with a target recall total of 500 items. The system first executes the high-priority recall path triggered by user behavior, obtaining the user's behavior sequence for the past three days. Based on this behavior sequence, it determines multiple recall levels and their priority order corresponding to the behavior-triggered recall path. Assuming the user behavior sequence includes clicking on product A (dress), adding product B (dress) to the cart, and clicking on product C (women's shoes), and the system predicts the target object's category is "dress," then based on the relevance of each behavior object's category to the "dress" category, products A and B (dress) are placed in the higher level, while product C (women's shoes) is placed in the lower level. The higher level is executed first for object recall. The system executes each level according to its priority order, adding the recalled objects from each level to the recall object set. Afterward, the system accumulates the number of recalled objects and determines whether the target recall total of 500 items has been reached. If the high-level recall has retrieved 300 but less than 500, the medium-priority recall path based on category matching will continue to supplement the recall; if the cumulative number reaches 500, the subsequent low-priority paths will be discontinued. Ultimately, the candidate objects retrieved by all recall paths constitute the final recall object set.

[0064] Based on this, by sequentially recalling objects according to multiple recall paths based on path priority information, differentiated processing of different recall paths is achieved. This ensures that recall results from high-priority recall paths are prioritized for inclusion in the recall object set, while low-priority recall paths are only used as supplements when the number of recalled objects is insufficient. This avoids low-quality recall results crowding out the recall space and improves the overall efficiency of the recall phase. Simultaneously, by determining multiple recall levels based on user behavior sequences within the behavior-triggered recall path and sequentially recalling objects according to the priority of these levels, fine-grained control within the same recall path is achieved. This prioritizes the recall of behavioral objects with higher semantic relevance to the recall request within the user behavior sequence, effectively controlling semantic drift and improving the relevance between recall results and the recall request. The synergistic effect of these two levels of hierarchical recall constructs a complete quality control system from both the dimensions between and within recall paths, reducing the number of low-quality candidate objects entering subsequent sorting stages and reducing ineffective computational overhead.

[0065] Furthermore, based on the path priority information, object recall is performed sequentially according to multiple recall paths to obtain a recall object set corresponding to the recall request, including: determining the recall path corresponding to the current path priority among the multiple recall paths based on the path priority information; performing object recall according to the recall path corresponding to the current path priority, and accumulating the recall quantity of the first recalled objects; if the recall quantity of the first recalled objects is less than the first recall quantity of the recall request, determining the recall path corresponding to the next path priority among the multiple recall paths based on the path priority information, and performing object recall according to the recall path corresponding to the next path priority; if the recall quantity of the first recalled objects is greater than or equal to the first recall quantity of the recall request, constructing a recall object set based on the first recalled objects.

[0066] The current path priority refers to the priority level of the currently executing recall path. The system executes each recall path sequentially from highest to lowest priority, and the current path priority determines which path should be executed at the moment. The next path priority refers to the next priority level after the current path priority. After the recall path corresponding to the current priority is completed, if the number of recalled objects is insufficient, the recall path corresponding to the next path priority will continue to be executed. The next path priority has a lower priority than the current path priority. The number of objects recalled in the first phase refers to the total number of objects recalled after executing the recall paths in order of path priority. The first recall number refers to the target total number of objects to be recalled in the recall request, that is, the total number of objects expected to be recalled during the recall phase. When the cumulative number of recalled objects reaches the target total number of objects, the recall ends.

[0067] In practical applications, the system recalls objects sequentially according to multiple recall paths based on path priority information. However, how to specifically execute the "sequential" process—that is, how to switch and judge between different priority paths, and how to determine when to stop executing the next path—is a specific problem that needs to be solved. Without clear switching conditions and a stopping mechanism, it's possible to waste computation by continuing to execute low-priority paths even when the number of recalled objects is sufficient, or to stop prematurely when the number of recalled objects is insufficient, resulting in inadequate recall coverage. Therefore, the embodiments in this specification further define the recall process as executing the recall according to the recall path corresponding to the current path priority and accumulating the number of recalled objects. When the number of recalled objects is less than a first recall quantity, the system continues to execute the next path priority. When the number of recalled objects reaches or exceeds the first recall quantity, the system stops and constructs a set of recalled objects, thereby achieving orderly switching and adaptive stopping between paths.

[0068] In practical implementation, as an optional approach, if the next path priority is the last path priority (i.e., the lowest priority), after executing the object recall according to the recall path corresponding to that last path priority, if the number of recalled objects is still less than the first recall quantity, execution will not continue because there are no lower priority paths available for execution. Instead, a recall object set will be directly constructed based on the recalled objects. This mechanism ensures that even without sufficient recall results, the system can still output the recalled objects as recall results, avoiding the occurrence of an empty recall result.

[0069] Furthermore, the specific value of the initial recall quantity can be flexibly set according to actual project needs, such as 5,000, 50,000, or any other suitable value. Different application scenarios can set different target recall quantities. For example, for scenarios requiring a large number of candidates for fine-grained sorting, the target recall quantity can be set larger; for scenarios prioritizing rapid response, the target recall quantity can be set smaller to reduce the processing load in subsequent sorting stages. The specific value of the initial recall quantity mentioned above can be dynamically adjusted according to project needs, and this specification does not impose any limitations on it.

[0070] In a specific embodiment of this specification, when a user searches for "red dress," the system determines multiple recall paths and their priorities as follows: user behavior-triggered recall paths are high priority, category matching-based recall paths are medium priority, and text matching-based recall paths are low priority. The initial recall quantity is 500 items. The system first determines the current path priority as high priority based on the path priority information, corresponding to the user behavior-triggered recall path. After executing this path, 300 candidate products are recalled, bringing the total number of recalled items to 300. Since 300 is less than 500, this indicates that the recall results from the high-priority recall path alone have not reached the target total recall quantity. Therefore, the system determines the next path priority as medium priority based on the path priority information, corresponding to the category matching-based recall path. This path is then executed to recall 200 candidate products, bringing the total number of recalled items to 500. At this point, the recalled quantity of 500 equals the initial recall quantity of 500. The system constructs a recall object set based on the recalled objects and no longer executes the low-priority text matching-based recall path.

[0071] For example, if a high-priority recall path triggered by user behavior recalls 600 candidate products, bringing the total number of recalled products to 600, and since 600 is greater than 500 and the target total number of recalls has been met, the system will directly construct a recall object set based on these 600 recalled objects and will no longer execute medium-priority and low-priority recall paths.

[0072] Based on this, the recall path corresponding to the current priority is determined and executed step by step according to the path priority information. After each execution, the number of recalled objects is accumulated and it is judged whether it is less than the first recall number. When the number of recalled objects is insufficient, the next priority path is executed. When the number of recalled objects reaches or exceeds the target, the construction of the recall object set is stopped. This realizes the adaptive switching and orderly connection between paths. It can not only ensure that the recall coverage can be automatically supplemented by low-level paths when the recall of high-level objects is insufficient, but also avoid the invalid computing overhead caused by continuing to execute low-priority paths when the target recall total has been met, thus improving the execution efficiency of the recall phase.

[0073] Furthermore, the next path priority is the last path priority; after executing object recall according to the recall path corresponding to the next path priority, the method further includes: accumulating the recall quantity of the first recalled objects; when the recall quantity of the first recalled objects is less than the first recall quantity, and the current path priority is the last path priority, constructing a recall object set based on the first recalled objects.

[0074] Here, "last path priority" refers to the lowest priority level among multiple recall paths. When the recall path corresponding to the last path priority is executed, it indicates that all available recall paths have been completed. "First recall quantity" refers to the target total number of objects to be recalled in the recall request, i.e., the total number of objects expected to be recalled during the recall phase. "First recalled object quantity" refers to the total number of objects recalled after executing the last path priority.

[0075] In practical applications, during the sequential execution of each recall path according to its priority, an edge case exists: after executing all available recall paths, the cumulative number of recalled objects still falls short of the target recall total. If the system continues to wait or attempt to execute non-existent paths at this point, the recall process cannot terminate normally; if it directly returns an empty result, it will affect the input quality and user experience of subsequent sorting stages. Therefore, in this embodiment, by determining whether the current path priority is the priority of the last path, if the cumulative recall number is still less than the first recall number after executing the last path, a recall object set is directly constructed based on the existing recalled objects, thereby ensuring that the recall process can terminate normally and output results after all paths have been executed.

[0076] In practice, the initial recall quantity is the target total recall quantity. If the recalled quantity is still less than the target value after all available recall paths have been executed, it is reasonable for the system to use the existing results as the final recall set, because there are no more available recall paths to execute, and continuing to wait or try invalid paths is meaningless. This mechanism ensures that the recall process has clear termination conditions and output results under any circumstances.

[0077] In a specific embodiment of this specification, when a user searches for "red dress," the system determines multiple recall paths and their priorities as follows: recall paths triggered by user behavior are high priority, recall paths based on category matching are medium priority, and recall paths based on text matching are low priority (i.e., last path priority). The initial recall quantity is 500. The system sequentially executes the high-priority path to recall 150 items, the medium-priority path to recall 120 items, and the cumulative recalled quantity is 270, still less than 500. At this point, the system determines the next path priority as low priority (i.e., last path priority) and continues to execute this last path to recall 80 items, bringing the cumulative recalled quantity to 350. Since all paths have been executed, the recalled quantity of 350 is still less than the initial recall quantity of 500, and the currently executed path has the last path priority, the system no longer waits or performs any further operations. It directly constructs a recall object set based on the 350 recalled candidate objects for subsequent sorting processing.

[0078] Based on this, after executing object recall according to the recall path corresponding to the last path priority, the number of recalled objects is accumulated. If the number of recalled objects is still less than the first recall number and the current path priority is the last path priority, the recall object set is directly constructed based on the recalled objects. This avoids the problem that the recall process cannot end normally or returns empty results because the number of recalled objects is still insufficient after all paths have been executed. This ensures that the recall process can terminate normally and output the existing recall results in all cases, thus improving the robustness and availability of the recall method.

[0079] Further, the process of executing object recall according to the recall path corresponding to the current path priority and accumulating the recall quantity of the first recalled objects includes: determining a preset path recall quantity for the recall path corresponding to the current path priority; executing object recall according to the recall path corresponding to the current path priority to obtain the first initial recalled objects for the recall path corresponding to the current path priority; calculating a first object efficiency score for the first initial recalled objects and sorting the first initial recalled objects according to the first object efficiency score to obtain sorted first initial recalled objects; if the recall quantity of the first initial recalled objects is greater than the preset path recall quantity, selecting objects from the sorted first initial recalled objects according to the sorting order and using the selected first initial recalled objects as the first recalled objects; if the recall quantity of the first initial recalled objects is less than the preset path recall quantity, using the first initial recalled objects as the first recalled objects.

[0080] The preset path recall quantity refers to the maximum number of recallable objects that the current path priority corresponds to when performing object recall. In other words, it is the maximum number of candidate objects that this path can contribute to the recall object set. The preset path recall quantity is used to control the scale of the recall results of a single recall path and prevent a path from crowding out the contribution space of other paths due to an excessive number of recalls.

[0081] The first initial recall objects refer to all initial candidate objects recalled along the recall path corresponding to the current path priority, without any sorting or truncation. The first object efficiency score is a quantitative score used to measure the quality or contribution of a single object in the recall results. The object efficiency score is used to sort the initial recall objects within the same recall path to determine which objects are preferentially retained as the recall results for that path. The object efficiency score can be determined based on a single indicator such as click-through rate, conversion rate, or popularity, or it can be determined based on a combination of multiple indicators. The sorted first initial recall objects refer to the ordered sequence of objects obtained after sorting the first initial recall objects according to their first object efficiency scores. In the sorted sequence, objects with higher object efficiency scores are ranked first, and objects with lower scores are ranked last.

[0082] In practical applications, when recalling objects according to the recall path corresponding to the current path priority, the number of initial candidate objects recalled by that path may exceed the preset upper limit of the number that can be contributed, and the quality of these initially recalled objects is uneven. If all initially recalled objects are directly used as recalled objects for that path, it may result in that path contributing too many candidate objects, crowding out the contribution space of other paths. At the same time, low-quality candidate objects may also enter the recalled object set, increasing the invalid computational overhead of the subsequent sorting stage. To address this, in the embodiments of this specification, object efficiency scores are calculated for the initial recalled objects and sorted. When the number of initial recalled objects exceeds the preset recall limit for a path, objects with high scores are selected as recalled objects according to the sorting order. Otherwise, all are retained, thereby achieving quality and quantity control of the recalled candidate objects within each recall path.

[0083] In practice, the efficiency score of the first object can be determined based on the static score of the candidate object. The static score is an efficiency score pre-calculated for each object, which can reflect the object's historical performance, such as historical click-through rate, historical conversion rate, object popularity, etc. The static score is pre-calculated based on the object's historical performance data and stored in the index library. For example, it can be calculated using the following formula (1): Formula (1) Wherein, Score(i) represents the efficiency score of object i; CTR(i) represents the historical click-through rate of object i, which is calculated as the ratio of the number of clicks to the number of exposures of object i in the historical exposure; CVR(i) represents the historical conversion rate of object i, which is calculated as the ratio of the number of times object i was converted after a historical click to the number of clicks; Pop(i) represents the normalized popularity value of object i, which can be obtained by normalizing the exposure or click volume of object i in the past preset time period; α, β, and γ are preset weight coefficients, satisfying α+β+γ=1, and each weight coefficient can be dynamically adjusted according to the target.

[0084] Objects recalled through different recall paths can be sorted using the same efficiency scoring system to ensure consistent object quality across different paths. When the initial number of recalled objects exceeds the preset recall quantity for each path, the system selects the same number of objects as the preset recall quantity from the sorted initial recall objects, in descending order of efficiency score, and discards the remaining objects. When the initial number of recalled objects is less than or equal to the preset recall quantity for each path, the system does not truncate the recall results for that path and retains all of them as recalled objects. This sorting and truncation process ensures that each recall path can contribute the highest quality candidate objects within its preset quantity range.

[0085] In a specific embodiment of this specification, the current path has a high priority, corresponding to a recall path triggered by user behavior, and the preset recall quantity for this path is 200. After the system executes this path, it obtains a first initial recall object of 300 candidate products. The system calculates the object efficiency score of each of these 300 candidate products, for example, by calculating a comprehensive score based on the historical click-through rate and conversion rate of each product, and sorts them from high to low scores. Since the initial recall object quantity of 300 is greater than the preset path recall quantity of 200, the system selects the first 200 products from the sorted 300 candidate products as the first recalled object of this path according to the sorting order, and the remaining 100 products with lower scores are truncated and do not enter the recall object set. If only 180 candidate products are obtained after the path is executed, and the initial recall object quantity of 180 is less than the preset path recall quantity of 200, then the system uses all 180 candidate products as the first recalled object.

[0086] Based on this, by determining the preset recall quantity corresponding to the current path, and calculating and sorting the object efficiency score of the initial recalled objects after executing the path recall, when the number of initial recalled objects exceeds the preset path recall quantity, objects with high scores are selected as recalled objects according to the sorting order. This not only achieves object quality screening within a single recall path, ensuring that candidates with high efficiency scores are retained first and low-quality candidates are truncated, but also achieves precise control over the number of candidate objects contributing to the path, avoiding the crowding out of the contribution space of other paths due to an excessive number of recalls in a single path, ensuring a reasonable allocation of recall scale among paths, and improving the overall quality of the recalled object set.

[0087] Furthermore, determining the user behavior sequence corresponding to the recall request includes: obtaining historical behavior data of the user corresponding to the recall request within a preset time period; sorting the historical behavior data by time to obtain the user behavior sequence corresponding to the recall request.

[0088] The preset time period refers to a pre-defined time range used to limit the time span for acquiring user historical behavior data. The preset time period ensures the timeliness of the acquired user behavior sequences, avoiding interference from outdated historical behaviors in the current recall. The preset time period can be flexibly set according to project needs, such as the last 3 days, the last 7 days, etc. Historical behavior data refers to the historical interaction records generated by users on the e-commerce platform, including but not limited to objects clicked, added to cart, purchased, and exposed. Historical behavior data also includes the time information corresponding to each behavior record, used for time-ordering of behaviors. Time-ordering refers to arranging historical behavior data according to the chronological order of the behaviors. In this step, time-ordering is used to organize the user's historical behavior data into an ordered sequence for subsequent selection of trigger objects. A user behavior sequence is a collection of user historical behavior data arranged in chronological order, with each behavior object in the sequence carrying corresponding behavior time information.

[0089] In practical applications, during the execution of the behavior-triggered recall path, it is necessary to determine the user behavior sequence corresponding to the recall request to provide a source of trigger objects for subsequent recall hierarchy division and object recall. However, users' historical behavior data is usually discrete and unordered. If these unorganized behavior data are used directly, it is impossible to distinguish the temporal relationship between different behavior objects, and it is also difficult to identify the user's recent interest change trend. To this end, in the embodiments of this specification, historical behavior data of users within a preset time period is obtained and sorted by time, organizing the discrete historical behavior into an ordered sequence, thereby providing a structured and time-series information-based data foundation for subsequently determining the recall hierarchy and trigger objects based on the user behavior sequence.

[0090] In practical implementation, regarding the choice of sorting direction, one option is to sort by behavior time from most recent to oldest, placing the most recently occurring behavior at the beginning of the user behavior sequence, making it easier to prioritize the most recent behavior as the trigger object in subsequent user behavior sequences. Another option is to sort by behavior time from oldest to most recent, which can be flexibly determined according to the needs of the actual recall strategy. The historical behavior data included in the user behavior sequence can include different types of behavior objects such as clicks, adding to cart, purchases, and impressions. The duration of the preset time period can be set according to project needs. For example, for users with sparse behavior, the preset time period can be appropriately extended to ensure sufficient behavior data is available, while for users with intensive behavior, the preset time period can be appropriately shortened to ensure data timeliness.

[0091] In a specific embodiment of this specification, a user's historical behavior data on an e-commerce platform includes multiple records, containing the user's historical interaction behavior within the last three days. For example: on day 1, the user clicked on product A; on day 1, the user added product B to their cart; on day 2, the user clicked on product C; on day 2, the user purchased product D; and on day 3, the user clicked on product E. The system obtains the aforementioned historical behavior data of the user within a preset time period and sorts it according to the order of the behavior occurrence from most recent to oldest. The resulting user behavior sequence is as follows: product E (clicked on day 3), product D (purchased on day 2), product C (clicked on day 2), product B (added to cart on day 1), and product A (clicked on day 1). This sorted sequence is the user behavior sequence corresponding to the recall request, serving as the basis for determining the triggering object and recall level in the subsequent behavior-triggered recall path.

[0092] Based on this, by acquiring users' historical behavior data within a preset time period and sorting the historical behavior data by time, disordered discrete historical behaviors can be transformed into ordered structured sequences. This allows the behavior-triggered recall path to distinguish between recent and distant user behaviors based on time sequence information during execution, thus prioritizing the recall of the user's most recent behavior as the trigger object. This makes the recall results more reflective of the user's current interests and preferences, improving the matching degree between the recall results and the user's real-time intent. At the same time, by limiting the preset time period, it avoids noise interference from overly outdated historical behaviors on the current recall decision.

[0093] Furthermore, determining multiple recall levels corresponding to the behavior-triggered recall path based on the user behavior sequence, and the priority order of the multiple recall levels, includes: determining the target object category corresponding to the recall request, and the object category of each behavior object in the user behavior sequence; dividing the object category of each behavior object according to the target object category to obtain multiple object category groups, wherein each object category group has a different degree of association with the target object category; determining multiple recall levels and the priority order of the multiple recall levels based on the multiple object category groups and the degree of association between each object category group and the target object category.

[0094] The target object category refers to the target category information corresponding to the recall request, that is, the category to which the object pointed to by the user's current query intent belongs. The target object category is used to compare with the categories of behavioral objects in the user's behavior sequence to determine the semantic relevance between each behavioral object and the current recall request. The target object category can be obtained by predicting the category from the text information contained in the recall request, or it can be determined based on the category identifier carried in the recall request. The object category refers to the category information in the classification system to which each behavioral object in the user's behavior sequence belongs. The object category is used to compare with the target object category to determine the degree of association between the behavioral object and the recall request. The object category group refers to different groupings obtained after dividing the behavioral objects in the user's behavior sequence according to the relationship between the object category of the behavioral object and the target object category. Different object category groups correspond to different degrees of association. The degree of association refers to the strength of the correlation between the object category of the behavioral object and the target object category. In the embodiments of this specification, the degree of association is used to determine the priority order of the recall levels corresponding to each object category group; the higher the degree of association, the higher the priority of the corresponding recall level.

[0095] In practical applications, within the behavior-triggered recall path, multiple recall levels and their priority order need to be determined based on the user behavior sequence. However, the categories of behavior objects in the user behavior sequence are diverse, and different behavior objects have varying degrees of semantic relevance to the current recall request. If all behavior objects are treated as trigger objects at the same level without distinction, it may lead to the triggering of recalls by behavior objects that are semantically unrelated to the recall request, resulting in semantic drift. To address this, the embodiments in this specification determine the target object category and the object category of each behavior object, divide the object category of each behavior object according to the target object category to obtain multiple object category groups with different degrees of relevance, and determine the corresponding recall level and priority order based on the degree of relevance between each object category group and the target object category, thereby achieving hierarchical division based on category relevance.

[0096] In practical implementation, the determination of the degree of association in the above classification method can adopt multiple granularities. As an example, behavioral objects whose object category and target object category are the same (i.e., belonging to the same leaf category) can be grouped into the group with the highest degree of association; behavioral objects whose object category and target object category belong to the same parent category (i.e., belonging to the same second-level category) can be grouped into the group with medium degree of association; and behavioral objects whose object category and target object category do not belong to the same parent category can be grouped into the group with the lowest degree of association. Through this multi-level category matching method, fine-grained grouping of behavioral objects in the user behavior sequence is achieved.

[0097] In practical implementation, if the number of behavior objects in a certain group of the user behavior sequence is 0, the recall level corresponding to that group can be skipped during actual execution, and the next level can be executed directly. For example, if there are no behavior objects in the user behavior sequence that match the target object category, the first group is empty, and the system can skip the first recall level and execute the second recall level directly. The target object category can be obtained by predicting the category from the text information in the recall request. When the recall request contains explicit search terms, the accuracy of category prediction is high. When the recall request does not contain text information, such as in similar recommendation scenarios, the target object category can be obtained directly from the recall request, for example, by obtaining the category information of the object currently viewed by the user as the target object category, or by obtaining the category identifier carried in the recall request. This specification does not limit this. In addition, the hierarchical division of the parent category and leaf category in the category system can be determined according to the structure of the actual category system. The leaf category and parent category mentioned in the embodiments of this specification are only examples. In application scenarios with multi-level category systems, other levels of categories can also be flexibly used as the basis for division. For example, in a deeply hierarchical category system, such as fourth or fifth level categories, behavioral objects belonging to the same third or fourth level category as the target object can be grouped into the same group, rather than being limited to leaf categories or first-level parent categories. The specific level used as the basis for division can be flexibly determined based on the actual depth of the category system and project requirements.

[0098] In a specific embodiment of this specification, when a user searches for "red dress," the system predicts the category of the search term and determines that the target object category is "dress." The user behavior sequence contains multiple behavior objects: Product E (category: dress), Product D (category: dress), Product C (category: women's shoes), Product B (category: dress, a subcategory of dress), and Product A (category: top). The system determines that the object category for each behavior object in the user behavior sequence is dress, dress, women's shoes, dress, and top, respectively. Based on the target object category "dress," the system categorizes the object categories of each behavior object: Products E, D, and B have object categories that are the same as or belong to a subcategory of "dress," and have the highest correlation with the target object category, thus belonging to the first object category group; Product C's object category, women's shoes, belongs to the same parent category (footwear and accessories) as "dress," with a medium correlation, and is belonging to the second object category group; Product A's object category, top, does not belong to the same parent category as "dress," with the lowest correlation, and is belonging to the third object category group. The system determines three recall levels based on the degree of association between the three object category groups and the target object category. The first object category group with the highest degree of association corresponds to the first recall level with the highest priority. The second object category group with a medium degree of association corresponds to the second recall level with the next lower priority. The third object category group with the lowest degree of association corresponds to the third recall level with the lowest priority.

[0099] Based on this, by determining the target object category corresponding to the recall request and the object category of each behavior object in the user behavior sequence, the object category of each behavior object is divided according to the target object category to obtain multiple object category groups with different degrees of relevance. Then, based on the degree of relevance between each object category group and the target object category, multiple recall levels and their priority order are determined. This achieves hierarchical division within the behavior-triggered recall path, so that behavior objects with a higher degree of semantic relevance to the current recall request are assigned to higher priority recall levels. Thus, during execution, recall is triggered with priority for behavior objects with high relevance, effectively controlling the cross-category semantic drift problem and improving the relevance between the recall results and the query intent.

[0100] Further, according to the hierarchical priority order, object recall is performed sequentially according to the multiple recall levels to obtain the recalled objects corresponding to the behavior-triggered recall path and add them to the recalled object set, including: determining the recall level corresponding to the current hierarchical priority among the multiple recall levels according to the hierarchical priority order; performing object recall according to the recall level corresponding to the current hierarchical priority and accumulating the recall quantity of the second recalled object; if the recall quantity of the second recalled object is less than the second recall quantity of the behavior-triggered recall path, determining the recall level corresponding to the next hierarchical priority among the multiple recall levels according to the hierarchical priority order and performing object recall according to the recall level corresponding to the next hierarchical priority; if the recall quantity of the second recalled object is greater than or equal to the second recall quantity of the behavior-triggered recall path, adding the second recalled object as the recalled object corresponding to the behavior-triggered recall path to the recalled object set.

[0101] The current priority level refers to the priority order of the currently executing recall level. The system executes each recall level sequentially from highest to lowest priority, with the current priority level determining which level of recall should be executed. The next priority level refers to the next priority tier below the current priority level. If the number of recalled objects is insufficient after the recall corresponding to the current level is completed, the recall level corresponding to the next priority level will continue to be executed. The priority of the next priority level is lower than that of the current priority level. The second recalled object count refers to the total number of objects recalled by the path after executing each recall level within the behavior-triggered recall path according to the priority order. The second recall count refers to the number of objects that the behavior-triggered recall path expects to contribute in this recall, i.e., the first maximum number of recallable objects preset for this path. The second recall count is used to control the overall recall scale of the behavior-triggered recall path. When the number of objects recalled by the path reaches this count, the execution of the next recall level within the path will stop.

[0102] In practical applications, after determining multiple recall levels and their priority order within the behavior-triggered recall path, specific problems need to be solved, such as how to execute the recall at each level, how to switch and judge between different levels, and how to determine when to stop executing the next level. Without a clear execution order and stopping mechanism, the following situations may occur: Executing a lower-priority level when a high-priority level has already recalled sufficient objects may result in lower-priority recalled objects irrelevant to the query intent being added to the recall results, causing semantic pollution; or failing to promptly activate the next level when a high-priority level has insufficient recall results may lead to insufficient recall quantity for that path, affecting recall coverage. Therefore, the embodiments in this specification further define the execution of each recall level sequentially from high to low priority, accumulating the recalled quantity. When the recalled quantity is less than the second recall quantity, execution continues to the next level; when the recalled quantity reaches or exceeds the second recall quantity, execution stops and the recall results for that path are output, thus achieving orderly switching and adaptive stopping between internal levels.

[0103] In practical implementation, as an optional approach, if the next level priority is the last level priority, after executing the object recall according to the recall level corresponding to that last level priority, if the number of the second recalled objects is still less than the second recall quantity, then execution will not continue, and the objects already recalled in that path will be directly added to the recall object set as the recall objects corresponding to the recall path triggered by this behavior. This mechanism ensures that even if the recall quantity is insufficient after all levels within that path have been executed, the recalled objects can still be output, avoiding the situation where no recall results are output for that path.

[0104] Furthermore, the specific value of the second recall quantity can be flexibly set according to the actual project needs. For example, it can be the same as the first recall quantity, or different values ​​can be allocated according to the expected contribution of each path. Each recall level can also be set with its own preset recall quantity during execution to control the truncation of recall results within each level.

[0105] In a specific embodiment of this specification, when a user searches for a red dress, the order of the multiple recall levels and their priorities within the behavior-triggered recall path is as follows: the first recall level (triggering behavior objects consistent with the dress category) has the highest priority, the second recall level (triggering behavior objects belonging to the same parent category as the dress category) has the next highest priority, and the third recall level (triggering behavior objects regardless of category) has the lowest priority. The second recall quantity for this path is 200. The system first determines the current level priority as the first recall level based on the level priority, executes the first recall level to recall objects, obtains candidate products recalled at this level, and accumulates 150 second recalled objects. At this point, 150 is less than 200, indicating that the recall at the first recall level alone has not yet reached the target contribution of this path. Therefore, the system determines the next level priority as the second recall level, continues to execute the second recall level to recall objects, obtains 80 candidate products recalled at this level, and accumulates 230 second recalled objects. At this point, 230 is greater than 200, which has reached the second recall quantity. The system stops executing the third recall level and adds the 230 candidate products that have already been recalled as the recall objects corresponding to the behavior-triggered recall path to the recall object set.

[0106] Based on this, by executing each recall level within the behavior-triggered recall path sequentially from high to low priority, and accumulating the recalled quantity after each execution and determining whether it is less than the second recall quantity, the system continues to execute the next level when the recalled quantity is insufficient, and stops when the recalled quantity reaches or exceeds the target quantity. This achieves adaptive switching and orderly connection between the levels within the behavior-triggered recall path. It ensures that when the recall of high-priority levels is insufficient, low-priority levels can be automatically used to supplement recall coverage, while avoiding the inclusion of objects unrelated to the query intent in the recall results when the target recall quantity of the path has been met. Thus, it effectively controls the risk of semantic drift while ensuring recall coverage.

[0107] Furthermore, the next-level priority is the last-level priority; after executing object recall according to the recall level corresponding to the next-level priority, the method further includes: accumulating the recall quantity of the second recalled object; if the recall quantity of the second recalled object is less than the second recall quantity of the behavior-triggered recall path, and the next-level priority is the last-level priority, then the second recalled object is added to the recall object set as the recall object corresponding to the behavior-triggered recall path.

[0108] The "last priority level" refers to the lowest priority level among the multiple recall levels within the behavior-triggered recall path. When the recall level corresponding to the last priority level is reached, it indicates that all available recall levels within that path have been executed. The "second recall quantity" refers to the number of objects that the behavior-triggered recall path expects to contribute in this recall, i.e., the maximum number of recallable objects preset for that path. The "second recalled object quantity" refers to the total number of objects recalled by the path after executing each recall level within the behavior-triggered recall path.

[0109] In practical applications, during the sequential execution of each recall level according to hierarchical priority within a behavior-triggered recall path, an edge case exists: after executing all available recall levels, the cumulative number of objects recalled for that path still falls short of the preset second recall quantity. If the system continues to wait or attempt to execute non-existent levels at this point, the recall process for that path cannot terminate normally; if the already recalled results are discarded directly, insufficient recall coverage will result. Therefore, in this embodiment, by determining whether the next level's priority is the last level's priority, if the cumulative recall quantity after executing the last level is still less than the second recall quantity, the existing second-recalled objects are directly added to the recall object set as the path's recall result, thereby ensuring that the recall process for that path can terminate normally and output the existing results after all levels have been executed.

[0110] In practice, the second recall quantity is the preset expected number of objects to contribute for that path. If the number of recalled objects is still less than this target value after all available recall levels have been executed, it is reasonable for the system to use the existing results as the recall output for that path, as there are no more available levels to execute, and continuing to wait or try invalid levels is meaningless. This mechanism ensures that behavior-triggered recall paths have clear termination conditions and output results under any circumstances. In extreme cases, if the number of second recalled objects is 0 after all recall levels have been executed (e.g., the user has no historical behavior data), the system will still use this empty result as the recall output for that path, and no objects from that path will be added to the recall object set. The overall recall process can still continue or terminate normally. Furthermore, the specific value of the second recall quantity can be flexibly set according to the path's historical contribution capacity and project needs. For example, a higher second recall quantity can be set for high-efficiency paths, and a lower second recall quantity can be set for low-efficiency paths.

[0111] In a specific embodiment of this specification, the order of multiple recall levels and their priorities within the behavior-triggered recall path is as follows: the first recall level has the highest priority, the second recall level is next, and the third recall level is the lowest priority. The second recall quantity for this path is 300. The system sequentially executes the first recall level to recall 150 objects, the second recall level to recall 80 objects, and the cumulative number of second recalled objects is 230, which is still less than 300. At this point, the system determines that the next level priority is the third recall level, i.e., the lowest priority, and continues to execute the recall of 40 objects at the lowest priority, bringing the cumulative number of second recalled objects to 270. At this point, all levels have been executed, the number of recalled objects (270) is still less than the second recall quantity (300), and the next level priority is the lowest priority. The system no longer waits or executes any further operations, and directly adds the 270 recalled objects as the recall objects corresponding to the behavior-triggered recall path to the recall object set.

[0112] Based on this, after executing object recall according to the recall level corresponding to the last level priority, the number of the second recalled objects is accumulated. If the number of recalled objects is still less than the number of the second recalled objects and the next level priority is the last level priority, the existing second recalled objects are directly added to the recall object set as the recall objects corresponding to the behavior-triggered recall path. This avoids the problem that the recall process cannot terminate normally or the existing recall results are lost because the number of recalled objects is still less than the target number after all levels in the path have been executed. This ensures that the behavior-triggered recall path can terminate normally and output the existing recall results in all cases, thus improving the robustness and availability of the recall method.

[0113] Further, the process of executing object recall according to the recall level corresponding to the current priority level and accumulating the recall quantity of the second recalled objects includes: determining a preset recall quantity for the recall level corresponding to the current priority level; executing object recall according to the recall level corresponding to the current priority level to obtain the second initial recalled objects corresponding to the recall level corresponding to the current priority level; calculating a second object efficiency score for the second initial recalled objects and sorting the second initial recalled objects according to the second object efficiency score to obtain sorted second initial recalled objects; if the recall quantity of the second initial recalled objects is greater than the preset recall quantity for the level, selecting objects from the sorted second initial recalled objects according to the sorting order and using the selected second initial recalled objects as the second recalled objects; if the recall quantity of the second initial recalled objects is less than the preset recall quantity for the level, using the second initial recalled objects as the second recalled objects.

[0114] The preset level recall quantity refers to the maximum number of recallable objects preset by the recall level corresponding to the current level priority when performing object recall. In other words, it's the maximum number of candidate objects that this level can contribute to the recall result triggered by the behavior. The preset level recall quantity is used to control the scale of the recall result for a single recall level, preventing one level from crowding out the contribution space of other levels due to an excessive number of recalls. The second initial recall objects refer to all initial candidate objects recalled by the current level priority after object recall is performed, without sorting or truncation. The second object efficiency score is a quantitative score used to measure the quality or contribution of a single object in the recall result. In the embodiments of this specification, the object efficiency score is used to sort the initial recall objects within the same recall level to determine which objects are preferentially retained as the recall result of that level. The second object efficiency score can use the same scoring system as the first object efficiency score to ensure consistency in object quality measurement standards across different levels and paths. The sorted second initial recall objects refer to the ordered sequence of objects obtained after sorting the second initial recall objects according to the second object efficiency score. In the sorted sequence, objects with higher object efficiency scores are ranked first, and objects with lower scores are ranked last.

[0115] In practical applications, when recalling objects according to the recall level corresponding to the current priority level within the behavior-triggered recall path, the number of initial candidate objects recalled at that level may exceed the preset upper limit of the number that can be contributed, and the quality of these initial recalled objects varies. If all initial recalled objects are directly used as recalled objects at that level, it may result in that level contributing too many candidate objects, crowding out the contribution space of other levels. At the same time, low-quality candidate objects may also enter the recalled object set, increasing the ineffective computational overhead of the subsequent sorting stage. To address this, the embodiments in this specification calculate object efficiency scores for the initial recalled objects and sort them. When the number of initial recalled objects exceeds the preset recall limit for each level, objects with high scores are selected as recalled objects according to the sorting order. Otherwise, all are retained, thereby achieving quality and quantity control of the recalled candidate objects within each recall level.

[0116] In practice, the efficiency score for the second object can use the same scoring system as the efficiency score for the first object, i.e., both use a unified static product efficiency score. This static score is an efficiency score pre-calculated for each object, reflecting its historical performance, such as historical click-through rate, historical conversion rate, and object popularity. Using the same efficiency scoring system to sort objects recalled at different recall levels ensures consistent object quality measurement standards across different levels, facilitating comparability of object quality during subsequent cross-level fusion. The calculation method for the efficiency score of the second object follows the same procedure as the efficiency score of the first object, and will not be elaborated further here.

[0117] When the number of initially recalled objects exceeds the preset recall quantity for each tier, the system selects the same number of objects as the preset recall quantity from the sorted initial recall objects, in descending order of object efficiency score. The remaining objects are truncated and discarded. The truncated objects do not participate in subsequent processing, reducing data volume while ensuring the quality of candidates entering the recall set. When the number of initially recalled objects is less than or equal to the preset recall quantity for each tier, the system does not truncate the recall results for that tier and retains all of them as recalled objects. This sorting and truncation process ensures that each recall tier can contribute the highest quality candidates within its preset quantity range. The preset recall quantities for each recall tier can be the same or different; for example, a larger preset recall quantity can be set for high-priority tiers, and a smaller preset recall quantity can be set for low-priority tiers, reflecting the differentiated contributions between different tiers.

[0118] In a specific embodiment of this specification, the current priority level within the behavior-triggered recall path is the first recall level, with a preset recall quantity of 150 items. After executing this level, the system obtains 200 candidate products as the second initial recall objects. The system calculates the second object efficiency score for each of these 200 candidate products, for example, by calculating a comprehensive score based on the historical click-through rate and conversion rate of each product, and sorts them from highest to lowest score. Since the initial recall object quantity of 200 is greater than the preset level recall quantity of 150, the system selects the top 150 products from the sorted 200 candidate products as the second recalled objects of this level, and the remaining 50 products with lower scores are truncated and do not enter the recall results of this path. If only 120 candidate products are obtained after executing this level, and the initial recall object quantity of 120 is less than the preset level recall quantity of 150, then the system uses all 120 candidate products as the second recalled objects.

[0119] Based on this, by determining the preset recall quantity corresponding to the current level, and calculating and sorting the object efficiency scores of the initial recalled objects after executing the level recall, when the number of initial recalled objects exceeds the preset level recall quantity, objects with high scores are selected as recalled objects according to the sorting order. This not only achieves object quality screening within a single recall level, ensuring that candidates with high efficiency scores are retained first and low-quality candidates are truncated, but also achieves precise control over the number of contributing candidates for that level, avoiding the crowding out of the contribution space of other levels due to an excessive number of recalls in a single level. This ensures a reasonable allocation of the recall scale among the levels within the behavior-triggered recall path and improves the overall quality of the recall results of the behavior-triggered recall path.

[0120] Furthermore, the method also includes: determining the object display order corresponding to the recalled object set based on the path priority information; and displaying the recalled objects in the recalled object set using the object display order.

[0121] The object display order refers to the order in which the recalled objects in the recall object set are displayed. This order is related to the path priority information of the recall paths from which each recalled object originates. The display order of the recalled objects can be determined according to the path priority of the recall paths from which they originate, from highest to lowest. Recalled objects from recall paths with higher path priority are displayed before those from recall paths with lower path priority. For example, objects recalled by recall paths with a first-priority (high-priority) path are displayed before objects recalled by recall paths with a second-priority (medium-priority) or third-priority (low-priority) path.

[0122] In practice, for multiple recalled objects originating from the same recall path priority, they are arranged in descending order of object efficiency score in the display order. The higher the object efficiency score, the earlier it is displayed. Based on determining the display order according to path priority, objects from the same priority source are further sorted according to efficiency score, so that high-quality objects within the same recall path receive a better display position.

[0123] In a specific embodiment of this specification, taking the path priority information including three priority levels as an example, the recall objects corresponding to the high priority level, such as the recall path triggered by user behavior, are arranged at the beginning of the display order; the recall objects corresponding to the medium priority level, such as the recall path based on category matching, are arranged in the middle of the display order; and the recall objects corresponding to the low priority level, such as the recall path based on text matching or the recall path based on store preference, are arranged at the end of the display order.

[0124] Furthermore, displaying the recalled objects in the recalled object set using the object display order specifically includes: in response to the recall result display instruction corresponding to the recall request, arranging and displaying the recalled objects in the recalled object set sequentially on the search results page according to the object display order. The display order is represented by the order in which the recalled objects are displayed on the search results page; recalled objects that are placed earlier in the display order are displayed before recalled objects that are placed later in the display order on the search results page.

[0125] In a specific embodiment of this specification, the client responds to the search command issued by the user and sends the recall request to the server; after the server performs object recall, it obtains a set of recalled objects and determines the object display order corresponding to the set of recalled objects in the manner described above before returning to the client; the client receives the recall result carrying object display order information and renders and displays each recalled object on the search page in the display order.

[0126] In another specific embodiment of this specification, the server returns the recall object set and the corresponding path priority information to the client, and the client determines the display order and displays the objects based on the path priority information.

[0127] Furthermore, when displaying the recall object set on the search results page, at least some of the objects in the recall object set may carry display identifiers corresponding to the path priority information. Display identifiers include sorting marks corresponding to the path priority of the recall path from which the recall object originated, such as text labels like "Recommended" or "Featured," or numerical serial numbers, and visual identifiers such as different colors, border styles, and background colors. Through these display identifiers, users can intuitively perceive the differences in path priority among different recall objects.

[0128] Based on this, the display order is determined by path priority information, so that high-quality recall objects with high path priority can be displayed in a prominent position on the search results page. Users can see the candidate objects that are more relevant to their search intent first, thereby improving the user's perceived relevance of the search results and the user experience.

[0129] In practical applications, the execution process of each step of the above object recall method can generate traceable system logs on the server side. Specifically, the system logs record at least the following information: multiple identified recall paths and their path priority information; the execution order, actual number of recalls, truncation threshold, and truncation result when executing each recall path; and the execution order, actual number of recalls, and truncation result when executing each recall level within the behavior-triggered recall path. The above log data is persistently stored in a structured form on the server-side storage medium and can be used to trace and verify the correctness and completeness of the recall process. The configuration data related to the recall path priority information and recall level information stored on the server side can be verified through code auditing or configuration auditing. The above configuration data is saved in a readable form in the server-side source code or configuration file, possessing objective verifiability, and can be used to review and confirm the implementation of the technical solution when needed. The above-mentioned log recording and configuration data storage method ensures that the technical solution described in the embodiments of this specification has traceable and verifiable technical characteristics during implementation, facilitating the provision of objective evidence in subsequent technical identification or review processes.

[0130] The following is in conjunction with the appendix Figure 3 Taking the object recall method provided in this specification as an example in the application of content recommendation, the object recall method will be further explained. Among other things, Figure 3 The flowchart of an object recall method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0131] Step 302: Determine multiple candidate recall paths and the historical hit data of each candidate recall path; sort the multiple candidate recall paths according to the historical hit data of each candidate recall path to obtain the sorted multiple candidate recall paths.

[0132] In one embodiment, the content recommendation platform acquires a full set of candidate recall paths, including recall paths based on user interest tags, recall paths based on content similarity, recall paths based on popular content, and recall paths based on social relationships. During the offline phase, the platform evaluates the historical hit rate of each path, calculating the proportion of content viewed by users recalled through historical recommendations for each path, i.e., the click-through rate (CTR). The evaluation results show that the CTR based on user interest tags is the highest, followed by the CTR based on content similarity, then the CTR based on popular content, and the CTR based on social relationships is the lowest. The platform then ranks the paths from highest to lowest CTR.

[0133] Step 304: Among the sorted candidate recall paths, select multiple recall paths according to the preset path priority number, and set the path priority information corresponding to each recall path.

[0134] In one embodiment, the platform presets three path priorities: high priority, medium priority, and low priority. Based on the ranking, the highest-ranking recall path (based on user interest tags) is set to high priority, the second-ranking path (based on content similarity) is set to medium priority, the third-ranking path (based on popular content) is set to low priority, and the remaining lower-ranking paths are temporarily disabled. This mapping relationship is configured as information that can be retrieved from an online lookup table.

[0135] Step 306: In response to the recall request, determine the multiple recall paths corresponding to the recall request and the path priority information of the multiple recall paths, wherein the multiple recall paths include at least the behavior-triggered recall path.

[0136] In one embodiment, when a user opens a content recommendation app and refreshes the recommendation stream, the platform receives the recall request and obtains the user's identifier information and the current request context. The system determines multiple recall paths corresponding to this request and the priority information of each path based on a pre-configured mapping relationship: recall paths based on user interest tags have high priority, recall paths based on content similarity have medium priority, and recall paths based on popular content have low priority. Among these, the recall path based on user interest tags is the behavior-triggered recall path, which uses content the user has previously viewed as the trigger object and recalls content based on the relationships between content.

[0137] Step 308: Determine the recall path corresponding to the current path priority among multiple recall paths based on the path priority information; execute object recall according to the recall path corresponding to the current path priority, and accumulate the recall quantity of the first recalled object.

[0138] In one embodiment, the target recall total is 300 items. The system first determines the current path priority as high priority based on path priority information, corresponding to the recall path based on user interest tags, and executes the recall of 150 candidate items for that path, accumulating the number of the first recalled objects to 150 items.

[0139] Step 310: If the number of recalled objects in the first recall is less than the number of objects recalled in the first recall request, determine the recall path corresponding to the next path priority among multiple recall paths based on the path priority information, and execute object recall according to the recall path corresponding to the next path priority.

[0140] In one embodiment, if the number of recalled items is 150, which is less than the target total of 300 items, the system determines the next path priority as medium priority based on the path priority information. The system then continues to execute the recall path based on content similarity to recall 120 candidate items, bringing the total number of recalled items to 270.

[0141] Optionally, after executing object recall according to the recall path corresponding to the next path priority, the recall quantity of the first recalled object is accumulated; if the recall quantity of the first recalled object is less than the first recall quantity and the current path priority is the last path priority, a recall object set is constructed based on the first recalled object.

[0142] In one embodiment, if the total number of recalled items after all three paths have been executed is only 240, which is still less than the target total of 300 items, and the currently executed path has a low priority, i.e., the last path priority, the system directly constructs a recall object set based on the 240 recalled candidate items and no longer waits.

[0143] Step 312: If the number of recalled objects in the first recalled object is greater than or equal to the number of objects recalled in the first recall request, construct a recall object set based on the first recalled object.

[0144] In one embodiment, if 150 high-priority paths and 180 medium-priority paths are recalled, and the total number of recalled paths reaches 330, which is greater than the target total of 300, the system directly constructs a recall object set based on the 330 recalled candidate contents and no longer executes the low-priority paths.

[0145] Optionally, determine the preset recall quantity of the recall path corresponding to the current path priority; perform object recall according to the recall path corresponding to the current path priority to obtain the first initial recalled object of the recall path corresponding to the current path priority; calculate the first object efficiency score of the first initial recalled object, and sort the first initial recalled object according to the first object efficiency score to obtain the sorted first initial recalled object; if the recall quantity of the first initial recalled object is greater than the preset path recall quantity, select from the sorted first initial recalled object according to the sorting order, and take the selected first initial recalled object as the first recalled object; if the recall quantity of the first initial recalled object is less than the preset path recall quantity, take the first initial recalled object as the first recalled object.

[0146] In one embodiment, the preset recall quantity for the user interest tag-based recall path is 200. After executing this path, 250 initial recall content items are obtained. The platform calculates the efficiency score of each item and sorts them from highest to lowest score. Since the initial recall quantity of 250 is greater than the preset path recall quantity of 200, the platform selects the top 200 high-scoring items from the sorted content as the recalled objects for this path, and the remaining 50 items are truncated. If the path only obtains 180 initial recall content items, which is less than the preset path recall quantity of 200, then all 180 items are considered as recalled objects.

[0147] Step 3081: Obtain the historical behavior data of the user corresponding to the recall request within a preset time period; sort the historical behavior data by time to obtain the user behavior sequence corresponding to the recall request.

[0148] In one embodiment, the platform obtains the user's historical behavior data for the past 7 days, including videos A, B, C, etc., that the user has watched, as well as the time information corresponding to each viewing record. The platform sorts the above historical behavior data in order of the behavior occurring from most recent to oldest, resulting in the following user behavior sequence: Video C (watched on day 7), Video B (watched on day 5), and Video A (watched on day 3). This sequence serves as the basis for subsequently determining the triggering object and the recall level.

[0149] Step 3082: Determine the target object category corresponding to the recall request, and the object category of each behavior object in the user behavior sequence; divide the object category of each behavior object according to the target object category to obtain multiple object category groups, wherein the degree of association between each object category group and the target object category is different; determine multiple recall levels and the priority order of the multiple recall levels according to the multiple object category groups and the degree of association between each object category group and the target object category.

[0150] In one embodiment, the platform predicts that the target object category currently requested by the user is technology content. The behavioral objects in the user's behavior sequence include: Video C (category: technology), Video B (category: technology), and Video A (category: comedy). The platform categorizes each behavioral object according to the target object category: Videos C and B, whose categories match the target object category, are grouped into the first object category group, with the highest correlation; Video A, categorized as comedy, belongs to the same parent category (general entertainment) as technology, and is grouped into the second object category group, with a moderate correlation. Based on the two object category groups and their correlation, the platform determines two corresponding recall levels: the first object category group corresponds to the first recall level with the highest priority, and the second object category group corresponds to the second recall level with the next lowest priority.

[0151] Step 3083: Determine the recall level corresponding to the current priority among multiple recall levels according to the priority order; execute object recall according to the recall level corresponding to the current priority, and accumulate the recall quantity of the second recalled object.

[0152] In one embodiment, the preset recall quantity for the behavior-triggered recall path is 200 items. The system determines the current level priority as the first recall level (technology category) according to the hierarchical priority order, and executes the recall of 150 candidate items at this level, bringing the total number of second recalled objects to 150.

[0153] Step 3084: If the number of recalled objects in the second recall is less than the number of objects in the second recall path triggered by the behavior, determine the next priority level in the multiple recall levels according to the hierarchical priority order, and execute object recall according to the next priority level.

[0154] In one embodiment, if the number of recalled items (150) is less than the preset recall number (200) for that path, the system determines the next level priority as the second recall level (general entertainment category), and continues to recall 80 candidate items at this level, bringing the total number of recalled items to 230.

[0155] Optionally, after recalling objects according to the recall level corresponding to the next level priority, the recall quantity of the second recalled object is accumulated; if the recall quantity of the second recalled object is less than the second recall quantity of the behavior-triggered recall path, and the next level priority is the last level priority, the second recalled object is added to the recall object set as the recall object corresponding to the behavior-triggered recall path.

[0156] In one embodiment, if the first level recalls 120 items, the second level recalls 50 items, and the third level (last level, regardless of category) recalls 30 items, the total number of items recalled after all levels are completed is only 200 items, which is still less than the preset recall number of 200 items for this path. Furthermore, if the next level has the priority of the last level, the system will directly add these 200 recalled items as the recall result of the recall path triggered by this behavior to the recall object set.

[0157] Step 3085: If the number of recalled objects of the second recalled object is greater than or equal to the number of recalled objects of the second recalled object of the behavior-triggered recall path, the second recalled object is added to the recall object set as the recalled object corresponding to the behavior-triggered recall path.

[0158] In one embodiment, if the first level recalls 150 items and the second level recalls 80 items, the cumulative number of recalled items reaches 230, which is greater than the preset recall number of 200 items for this path. The system stops executing the next level and adds the 230 recalled candidate items as the recall objects corresponding to the recall path triggered by this behavior to the recall object set.

[0159] Optionally, determine the preset recall quantity for the recall level corresponding to the current priority level; perform object recall according to the recall level corresponding to the current priority level to obtain the second initial recalled objects corresponding to the recall level corresponding to the current priority level; calculate the second object efficiency score of the second initial recalled objects, and sort the second initial recalled objects according to the second object efficiency score to obtain the sorted second initial recalled objects; if the recall quantity of the second initial recalled objects is greater than the preset recall quantity for the level, select from the sorted second initial recalled objects according to the sorting order, and take the selected second initial recalled objects as the second recalled objects; if the recall quantity of the second initial recalled objects is less than the preset recall quantity for the level, take the second initial recalled objects as the second recalled objects.

[0160] In one embodiment, the preset recall quantity for the first recall tier is 120 items. After executing this tier, 160 initial recall items are obtained. The platform calculates the efficiency score for each item and sorts them from highest to lowest score. Since the initial recall quantity of 160 is greater than the preset tier recall quantity of 120, the platform selects the top 120 high-scoring items from the sorted content as the recalled objects for this tier, and the remaining 40 items are truncated. If this tier only obtains 100 initial recall items, which is less than the preset tier recall quantity of 120, then all 100 items are considered as recalled objects.

[0161] Corresponding to the above method embodiments, this specification also provides embodiments of an object recall system. Figure 4 This diagram illustrates the structure of an object retrieval system according to an embodiment of this specification. The system 400 includes a client 402 and a server 404. The client, in response to a user's search command on a search page, generates a retrieval request and sends it to the server. The server executes object retrieval according to the object retrieval method, obtains a set of retrieved objects corresponding to the retrieval request, and returns it to the client. The client displays the retrieved objects corresponding to the retrieved object set on the search page.

[0162] In this context, "client" refers to an application or webpage front-end running on the user's terminal device, used to provide a user interface, receive user commands, and display search results. The client can take various forms, such as a mobile app, a PC webpage, or a mini-program; this specification does not limit this. "Search command" refers to the command generated when a user enters a search term or triggers a search operation on the client's search page, instructing the system to begin the search process. Search commands can contain text information entered by the user or information about the user's current browsing context. "Recall request" is a request message generated by the client in response to a search command, requesting the server to perform object recall. This request carries user identification information, search term information, and request context information. The recall request is the data carrier for recall interaction between the client and the server.

[0163] The server refers to the backend service system deployed on the cloud or local server, which is used to receive recall requests from clients, execute the object recall methods described in the foregoing embodiments of this specification, and return the recall results to the client.

[0164] In practical applications, the server needs to execute complex object retrieval methods, including determining multiple retrieval paths, prioritizing paths, defining multi-level divisions within the behavior-triggered retrieval paths, and dynamic execution logic. The client needs to provide a user interface and result display functionality. To address this, this system generates a retrieval request in response to a user's search command and sends it to the server. The server executes the object retrieval method to obtain a set of retrieved objects and returns it to the client. The client then displays the retrieved objects on the search page. This clarifies the interaction flow and division of responsibilities between the client and server, providing system architecture support for the specific implementation of the aforementioned method embodiments.

[0165] In a specific embodiment of this specification, a user of a content recommendation platform enters "sports shoes" in the search box of an e-commerce app and clicks the search button. The client responds to this search instruction by generating a recall request, which includes user identification information and search term information. The client sends this recall request to the server via the network. After receiving the recall request, the server executes the object recall method described in the preceding embodiments: First, it determines multiple recall paths corresponding to this request and the path priority information of each path. For example, it determines that recall paths triggered by user behavior are of high priority, recall paths based on category matching are of medium priority, and recall paths based on text matching are of low priority. Then, it executes each path sequentially according to its priority to perform object recall. For behavior-triggered recall paths, it further determines the user's behavior sequence over the past 3 days, divides multiple recall levels according to the relevance between the category of the behavior object and the category of the target object, and executes each level of recall sequentially according to its priority. Finally, it obtains a set of recalled objects, such as 500 candidate products. The server returns the set of recalled items to the client, which then displays these 500 candidate products in a list format on the search results page for the user to browse and select.

[0166] Based on this, the client generates a recall request in response to the user's search command and sends it to the server. The server executes the recall according to the aforementioned object recall method and returns the recalled object set to the client. The client then displays the recalled objects on the search page, forming a complete client-server interaction loop. This provides system architecture support for the specific implementation of the aforementioned method embodiments. Furthermore, by centrally deploying the recall logic on the server, updates and optimizations of the recall strategy can take effect without upgrading the client, reducing system maintenance costs and improving the flexibility of recall strategy iteration. Through decoupling between the client and server, the same set of server-side recall logic can support clients on different platforms, exhibiting good cross-platform compatibility.

[0167] Corresponding to the above method embodiments, this specification also provides embodiments of an object recall device. Figure 5A schematic diagram of an object recall device according to one embodiment of this specification is shown. Figure 5 As shown, the device includes: The determination module 502 is configured to, in response to a recall request, determine multiple recall paths corresponding to the recall request and path priority information of the multiple recall paths, wherein the multiple recall paths include at least a behavior-triggered recall path. The module 504 is configured to sequentially recall objects according to multiple recall paths based on the path priority information, and obtain the set of recalled objects corresponding to the recall request. The process of recalling objects according to the behavior-triggered recall path includes: determining the user behavior sequence corresponding to the recall request, and determining multiple recall levels corresponding to the behavior-triggered recall path based on the user behavior sequence, as well as the priority order of the multiple recall levels; and sequentially recalling objects according to the priority order of the multiple recall levels to obtain the recalled objects corresponding to the behavior-triggered recall path and add them to the recall object set.

[0168] Optionally, the device further includes a setting module configured to determine multiple candidate recall paths and historical hit data for each candidate recall path; sort the multiple candidate recall paths according to the historical hit data of each candidate recall path to obtain multiple sorted candidate recall paths; select multiple recall paths from the multiple sorted candidate recall paths according to a preset path priority number, and set path priority information corresponding to each recall path.

[0169] Optionally, the obtaining module 504 is further configured to: determine the recall path corresponding to the current path priority among multiple recall paths based on the path priority information; perform object recall according to the recall path corresponding to the current path priority, and accumulate the recall quantity of the first recalled objects; if the recall quantity of the first recalled objects is less than the first recall quantity of the recall request, determine the recall path corresponding to the next path priority among the multiple recall paths based on the path priority information, and perform object recall according to the recall path corresponding to the next path priority; if the recall quantity of the first recalled objects is greater than or equal to the first recall quantity of the recall request, construct a recall object set based on the first recalled objects.

[0170] Optionally, the obtaining module 504 is further configured to accumulate the number of recalls of the first recalled objects; and when the number of recalls of the first recalled objects is less than the first recall number, and the current path priority is the end path priority, construct a recall object set based on the first recalled objects.

[0171] Optionally, the obtaining module 504 is further configured to: determine the preset path recall quantity of the recall path corresponding to the current path priority; perform object recall according to the recall path corresponding to the current path priority to obtain the first initial recall object of the recall path corresponding to the current path priority; calculate the first object efficiency score of the first initial recall object, and sort the first initial recall object according to the first object efficiency score to obtain the sorted first initial recall object; if the recall quantity of the first initial recall object is greater than the preset path recall quantity, select from the sorted first initial recall object according to the sorting order, and take the selected first initial recall object as the first recalled object; if the recall quantity of the first initial recall object is less than the preset path recall quantity, take the first initial recall object as the first recalled object.

[0172] Optionally, the obtaining module 504 is further configured to obtain historical behavior data of the user corresponding to the recall request within a preset time period; sort the historical behavior data by time to obtain the user behavior sequence corresponding to the recall request.

[0173] Optionally, the obtaining module 504 is further configured to determine the target object category corresponding to the recall request, and the object category of each behavior object in the user behavior sequence; divide the object category of each behavior object according to the target object category to obtain multiple object category groups, wherein each object category group has a different degree of association with the target object category; and determine multiple recall levels and the hierarchical priority order of the multiple recall levels according to the multiple object category groups and the degree of association between each object category group and the target object category.

[0174] Optionally, the obtaining module 504 is further configured to: determine the recall level corresponding to the current priority among the plurality of recall levels according to the hierarchical priority order; perform object recall according to the recall level corresponding to the current priority, and accumulate the recall quantity of the second recalled object; if the recall quantity of the second recalled object is less than the second recall quantity of the behavior-triggered recall path, determine the recall level corresponding to the next priority among the plurality of recall levels according to the hierarchical priority order, and perform object recall according to the recall level corresponding to the next priority; if the recall quantity of the second recalled object is greater than or equal to the second recall quantity of the behavior-triggered recall path, add the second recalled object as the recall object corresponding to the behavior-triggered recall path to the recall object set.

[0175] Optionally, the obtaining module 504 is further configured to accumulate the number of recalls of the second recalled object; if the number of recalls of the second recalled object is less than the number of recalls of the behavior-triggered recall path, and the next level priority is the last level priority, the second recalled object is added to the recall object set as the recall object corresponding to the behavior-triggered recall path.

[0176] Optionally, the obtaining module 504 is further configured to: determine a preset recall quantity for the recall level corresponding to the current priority level; perform object recall according to the recall level corresponding to the current priority level to obtain a second initial recall object corresponding to the recall level corresponding to the current priority level; calculate a second object efficiency score for the second initial recall object, and sort the second initial recall object according to the second object efficiency score to obtain a sorted second initial recall object; if the recall quantity of the second initial recall object is greater than the preset recall quantity for the current priority level, select from the sorted second initial recall object according to the sorting order, and take the selected second initial recall object as the second recalled object; if the recall quantity of the second initial recall object is less than the preset recall quantity for the current priority level, take the second initial recall object as the second recalled object.

[0177] Optionally, the device further includes a display module configured to determine the object display order corresponding to the recalled object set based on the path priority information; and to display the recalled objects in the recalled object set using the object display order.

[0178] The above is a schematic scheme of an object recall device according to this embodiment. It should be noted that the technical solution of this object recall device and the technical solution of the object recall method described above belong to the same concept. For details not described in detail in the technical solution of the object recall device, please refer to the description of the technical solution of the object recall method described above.

[0179] Figure 6 A structural block diagram of a computing device 600 according to one embodiment of this specification is shown. The components of the computing device 600 include, but are not limited to, a memory 610 and a processor 620. The processor 620 is connected to the memory 610 via a bus 630, and a database 650 is used to store data.

[0180] The computing device 600 also includes an access device 640, which enables the computing device 600 to communicate via one or more networks 660. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 640 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0181] In one embodiment of this specification, the above-described components of the computing device 600 and Figure 6 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 6 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0182] The computing device 600 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 600 can also be a mobile or stationary server.

[0183] The processor 620 is configured to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the object recall method described above.

[0184] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the object recall method described above belong to the same concept. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the object recall method described above.

[0185] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the object recall method described above.

[0186] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium and the technical solution of the object retrieval method described above belong to the same concept. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the object retrieval method described above.

[0187] An embodiment of this specification also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the object recall method described above.

[0188] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product and the technical solution of the object recall method described above belong to the same concept. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solution of the object recall method described above.

[0189] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0190] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0191] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0192] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0193] The preferred embodiments disclosed above are merely illustrative of this specification. Optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described in this specification. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification.

Claims

1. An object recall method, comprising: In response to a recall request, multiple recall paths corresponding to the recall request and path priority information of the multiple recall paths are determined, wherein the multiple recall paths include at least a behavior-triggered recall path; Based on the path priority information, objects are recalled sequentially according to multiple recall paths to obtain the set of recalled objects corresponding to the recall request; The process of recalling objects according to the behavior-triggered recall path includes: determining the user behavior sequence corresponding to the recall request, and determining multiple recall levels corresponding to the behavior-triggered recall path based on the user behavior sequence, as well as the priority order of the multiple recall levels; and sequentially recalling objects according to the priority order of the multiple recall levels to obtain the recalled objects corresponding to the behavior-triggered recall path and add them to the recall object set.

2. The method of claim 1, further comprising, in response to a recall request: Identify multiple candidate recall paths and the historical hit data for each candidate recall path; Based on the historical hit data of each candidate recall path, the multiple candidate recall paths are sorted to obtain multiple sorted candidate recall paths. Among the sorted candidate recall paths, multiple recall paths are selected according to a preset path priority number, and path priority information is set for each recall path.

3. The method according to claim 1, wherein object recall is performed sequentially according to multiple recall paths based on the path priority information to obtain a set of recalled objects corresponding to the recall request, comprising: Based on the path priority information, determine the recall path corresponding to the current path priority among multiple recall paths; Execute object recall according to the recall path corresponding to the current path priority, and accumulate the recall count of the first recalled object; If the number of recalled objects in the first recalled object is less than the number of recalled objects in the first recall request, the recall path corresponding to the next path priority is determined from the plurality of recall paths according to the path priority information, and the object recall is performed according to the recall path corresponding to the next path priority. If the number of recalled objects in the first recalled object is greater than or equal to the first number of recalled objects in the recall request, a recall object set is constructed based on the first recalled object.

4. The method according to claim 3, wherein the next path priority is the end path priority; After retrieving objects according to the recall path corresponding to the next path priority, the method further includes: The cumulative number of recalled entities that were initially recalled; If the number of recalled objects of the first recalled object is less than the number of recalled objects, and the current path priority is the end path priority, a set of recalled objects is constructed based on the first recalled object.

5. The method according to claim 3, wherein object recall is performed according to the recall path corresponding to the current path priority, and the recall quantity of the first recalled objects is accumulated, comprising: Determine the preset path recall quantity for the recall path corresponding to the current path priority; Execute object recall according to the recall path corresponding to the current path priority, and obtain the first initial recall object of the recall path corresponding to the current path priority; Calculate the first object efficiency score of the first initial recall object, and sort the first initial recall objects according to the first object efficiency score to obtain the sorted first initial recall objects; If the number of recalled objects in the first initial recall is greater than the number of recalled objects in the preset path, objects are selected from the sorted first initial recall objects according to the sorting order, and the selected first initial recall objects are taken as the first recalled objects. If the number of recalled objects in the first initial recall is less than the number of recalled objects in the preset path, the first initial recall object will be regarded as the first recalled object.

6. The method according to claim 1, wherein determining the user behavior sequence corresponding to the recall request includes: Obtain the historical behavior data of the user corresponding to the recall request within a preset time period; The historical behavior data is sorted by time to obtain the user behavior sequence corresponding to the recall request.

7. The method according to claim 1, wherein determining multiple recall levels corresponding to the behavior-triggered recall path based on the user behavior sequence, and the priority order of the multiple recall levels, includes: Determine the target object category corresponding to the recall request, and the object category of each behavior object in the user behavior sequence; Based on the target object category, the object category of each behavior object is divided to obtain multiple object category groups, wherein each object category group has a different degree of association with the target object category; Based on the multiple object category groups and the degree of association between each object category group and the target object category, multiple recall levels and the priority order of the multiple recall levels are determined.

8. The method according to claim 1, wherein objects are recalled sequentially according to the multiple recall levels based on the hierarchical priority order, and the recalled objects corresponding to the behavior-triggered recall path are obtained and added to the recall object set, comprising: The recall level corresponding to the current priority level is determined among the multiple recall levels according to the priority order of the levels; The object recall is performed according to the recall level corresponding to the current priority level, and the recall quantity of the second recalled object is accumulated; If the number of recalled objects in the second recall is less than the number of recalled objects in the second recall path triggered by the behavior, the recall level corresponding to the next level priority is determined in the multiple recall levels according to the hierarchical priority order, and the object recall is performed according to the recall level corresponding to the next level priority. If the number of recalled objects of the second recalled object is greater than or equal to the number of recalled objects of the behavior-triggered recall path, the second recalled object is added to the recall object set as the recall object corresponding to the behavior-triggered recall path.

9. The method according to claim 8, wherein the next level priority is the last level priority; After retrieving objects according to the recall level corresponding to the next lower priority level, the method further includes: The cumulative number of recalled entities that were previously recalled; If the number of recalled objects of the second recalled object is less than the number of recalled objects of the behavior-triggered recall path, and the next level priority is the last level priority, the second recalled object is added to the recall object set as the recall object corresponding to the behavior-triggered recall path.

10. The method according to claim 8, wherein object recall is performed according to the recall level corresponding to the current priority level, and the recall quantity of the second recalled objects is accumulated, comprising: Determine the preset recall quantity for the recall level corresponding to the current priority level; The object is recalled according to the recall level corresponding to the current priority level, and the second initial recall object corresponding to the recall level corresponding to the current priority level is obtained; Calculate the second object efficiency score of the second initial recall object, and sort the second initial recall objects according to the second object efficiency score to obtain the sorted second initial recall objects; If the number of recalled objects in the second initial recall is greater than the number of recalled objects in the preset level, objects are selected from the sorted second initial recall objects according to the sorting order, and the selected second initial recall objects are used as the second recalled objects. If the number of recalled objects in the second initial recall is less than the number of recalled objects in the preset level, the second initial recall object will be regarded as the second recalled object.

11. The method according to any one of claims 1-10, wherein the method further comprises: The order in which the objects corresponding to the recalled object set are displayed is determined based on the path priority information. The recalled objects in the recalled object set are displayed in the order of the object display.

12. An object retrieval system, the system comprising a client and a server, wherein, The client, in response to the search command issued by the user for the search page, generates a recall request and sends it to the server; The server performs object recall according to any one of claims 1-11, obtains the recall object set corresponding to the recall request, and returns it to the client. The client displays the recalled objects corresponding to the recalled object set on the search page.

13. A computing device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the object recall method according to any one of claims 1 to 11.

14. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the object recall method according to any one of claims 1 to 11.

15. A computer program product comprising a computer program or instructions which, when executed by a processor, implement the steps of the object recall method according to any one of claims 1 to 11.

Images