A collaborative adaptive posterior enhancement long tail recommendation method and a recommendation system

Abstract

The application provides a kind of collaborative adaptive posterior enhancement long tail recommendation method and recommendation system, including the following steps: obtaining the candidate item list generated by constraint search of generative recommendation model for user and the log probability score of each candidate item;Based on the historical interaction sequence of user, the collaborative prior of each candidate item is calculated using the precomputed item co-occurrence matrix, and the modified score of long tail candidate item is obtained by log space posterior correction;According to the length of user's history, the modification strength of modified score is dynamically adjusted by adaptive gating function;Apply ranking condition intervention strategy, lock the original Top-1 prediction unchanged, only rearrange the candidate items ranked from 2 to K according to the modified score, and output the final recommendation list.This method does not depend on additional training data or optimization process, and is applied to various recommendation scenarios based on autoregressive semantic ID generation in the form of post-processing plug-in.

Description

Technical Field

[0001] This application relates to the field of recommender system technology, specifically a collaborative adaptive posterior augmentation long-tail recommendation method. Background Technology

[0002] Generative recommender systems are an important emerging paradigm in recommender systems in recent years. These methods represent each item as a sequence of semantic IDs composed of several discrete tokens, typically obtained through residual quantization of content embeddings. The user's interaction history is mapped to a token sequence, and the model autoregressively generates the semantic ID of the target item based on this history, thus enabling prediction of the next item. During inference, candidate items are generated through a progressive expansion using bundle search, retaining the candidate sequences with the highest cumulative log probability at each step, and finally outputting a sorted list of candidate items.

[0003] However, a deeper analysis of the candidate set generated by the bundle search reveals a prevalent failure mode: bundle-induced long-tail collapse. This failure manifests on two levels: first, the exclusion problem, where long-tail target items are pruned during the bundle search process, and their probability of appearing in the candidate set is much lower than that of head items; second, the ranking suppression problem, where even if long-tail items enter the candidate set, their scores are significantly suppressed due to the cumulative amplification of popularity bias in multiple decoding steps, resulting in a ranking far lower than that of head items.

[0004] The root cause of these two types of problems lies in the fact that head items dominate the training data, causing the generator to tend to assign them higher probabilities at each decoding step. Beam search, with its greedy pruning based on cumulative log probability, gradually amplifies the popularity bias at each step, creating a cascading amplification effect. This effect makes it impossible for debiasing strategies during training to fundamentally eliminate inference bias.

[0005] Existing methods primarily address the training phase, including resampling, inverse bias weighting, causal intervention, and counterfactual training. However, none of these methods can prevent bundle search from reintroducing and accumulating popularity bias during inference. A few inference-time diversification methods, such as diverse bundle search and maximum marginal relevance reordering, mainly target candidate diversity and do not address the long-tail score suppression problem. Experiments show that these methods may even further reduce the hit rate of long-tail items.

[0006] In summary, existing technologies lack a method specifically designed to address long-tail bias during generative recommendation bundle search inference, without requiring retraining of the backbone model, and to strictly maintain overall accuracy while restoring long-tail hit rate. Summary of the Invention

[0007] The problem addressed by this invention is the systematic bias of long-tail items introduced by bundle search decoding in generative recommendation systems.

[0008] To address the aforementioned problems, this invention provides a collaborative adaptive posterior augmentation long-tail recommendation method, a recommendation system, an electronic device, and a storage medium.

[0009] In a first aspect, the present invention provides a collaborative adaptive posterior-enhanced long-tail recommendation method, comprising the following steps: Obtain the list of candidate items generated by the generative recommendation model for the user through constraint bundle search, and the log probability score of each candidate item; Based on the user's historical interaction sequence, the collaborative prior of each candidate item is calculated using a pre-computed item co-occurrence matrix, and the corrected score of the long-tail candidate item is obtained through log space posterior correction. The correction strength of the correction score is dynamically adjusted based on the user's historical length using an adaptive gating function; By applying a ranking condition intervention strategy, the original Top-1 prediction of the locked bundle search remains unchanged, and only the candidate items ranked from 2nd to Kth are re-ranked according to the modified scores to output the final recommendation list; The corrected score is expressed as: ; in, Represents the logarithmic probability of the model. This is a normalized long-tail logarithm prior that applies only to long-tail candidate items. To correct the intensity hyperparameter.

[0010] Optionally, the method for constructing the collaborative prior includes: Construct a sparse item co-occurrence matrix from the training interaction data. ; For users Based on its historical interaction sequence Laplace smoothing was used to calculate the user-level co-occurrence mean. ,in Statistics on items and items The number of users who generated interaction; Normalize the mean of user-level co-occurrence to obtain the collaborative prior distribution: , ; in, Candidate items In users Normalized cooperative prior probabilities under the following conditions; For candidate items The average co-occurrence at the user level; The normalization constant is the mean co-occurrence value of all candidate items. Summation, making This constitutes a valid probability distribution.

[0011] Optionally, the normalized long-tailed logarithmic prior... Defined as: ; in, For long-tail indicator functions, Let the smallest non-zero collaborative prior value among all candidate items be such that... ≥0 and head items .

[0012] Optionally, the adaptive gating function is a saturated linear gating function: ; in, Candidate items In users Normalized cooperative prior probabilities under the following conditions; For candidate items The average co-occurrence at the user level; The normalization constant is the mean co-occurrence value of all candidate items. Summation, making This constitutes a valid probability distribution.

[0013] Optionally, the ranking condition intervention strategy specifically includes: Compute the difference Δ between the log probability scores of the original Top-1 and Top-2 candidates in the beam search. 12 (u); Calculate the maximum correction magnitude λ·h among all long-tail candidate items. max ; If λ·h max ≥Δ 12 If (u), then the final ranking of the original Top-1 candidate is locked as first place, and the corrected score is applied to the candidate items ranked 2nd to Kth in the original ranking for re-ranking. Otherwise, apply the corrected scores to all candidate items and rearrange them.

[0014] Optionally, the method does not modify any parameters of the generative recommendation model, does not rely on additional training data or optimization processes, and can be integrated into the inference process of an existing generative recommendation system as a post-processing plugin.

[0015] Optionally, it also includes an offline construction phase: collecting user interaction sequences, mapping items to semantic ID sequences, counting the frequency of occurrence of all item pairs, and constructing and storing a sparse item co-occurrence matrix.

[0016] Optionally, the generative recommendation model is an autoregressive generative model based on semantic IDs, which generates a sequence of semantic IDs for candidate items and their corresponding cumulative log probabilities by decoding through constraint bundle search.

[0017] Optionally, the long-tail items are defined as items whose interaction frequency ranks in the bottom 80% of the training data, and the head items are items that rank in the top 20%.

[0018] Secondly, embodiments of the present invention provide a collaborative adaptive posterior-enhanced long-tail recommendation system, comprising: The co-occurrence matrix storage module is used to store the sparse item co-occurrence matrix obtained in advance from the training interaction data; The collaborative prior calculation module is used to query the co-occurrence matrix based on the user's historical interaction sequence, calculate the user-level co-occurrence mean and normalized collaborative prior of each candidate item, and generate a normalized long-tail logarithmic prior that only applies to long-tail candidate items. The adaptive gating module is used to calculate the saturated linear gating coefficient based on the user's historical data and dynamically adjust the correction strength. The posterior correction module is used to add the logarithmic probability scores of candidate items output by the model to the gated long-tailed logarithmic prior in the logarithmic space to obtain the corrected rearrangement scores. The ranking condition intervention module is used to detect the difference between the original Top-1 and Top-2 scores in the beam search. When the correction magnitude may change the Top-1 prediction, the original Top-1 prediction is locked, and the corrected scores are applied only to the candidate items ranked 2nd to Kth to re-rank them, and the final recommendation list is output.

[0019] Thirdly, embodiments of the present invention provide an electronic device, including a processor, a communication interface, a memory, and a bus, wherein the processor, the communication interface, and the memory communicate with each other through the bus, and the processor can call logical instructions in the memory to execute the steps of the method provided in the first aspect.

[0020] Fourthly, embodiments of the present invention provide a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the collaborative adaptive posterior augmentation long-tail recommendation method as described in the first aspect.

[0021] The beneficial effects of the collaborative adaptive posterior augmentation long-tail recommendation method of the present invention are as follows: This method acquires a candidate list and log probabilities for bundle search, constructs a collaborative prior based on user history and a pre-computed co-occurrence matrix, performs posterior correction on long-tail candidates in log space, and dynamically adjusts the correction intensity based on the length of user history using adaptive gating. Simultaneously, a ranking condition intervention strategy locks the original Top-1 prediction unchanged, only re-ranking subsequent items. It can be deployed as a plug-and-play plugin without retraining or modifying backbone model parameters, fundamentally eliminating the problem of long-tail item exclusion and score suppression caused by the cascading amplification of popularity bias during bundle search decoding, thus accurately improving the hit rate of long-tail items. Adaptive gating effectively protects users with cold starts or short histories from unreliable prior interference, and the ranking condition intervention strategy strictly ensures that the overall Top-1 recommendation accuracy does not decrease, thereby significantly improving long-tail recommendation coverage while maintaining or even improving overall recommendation performance. Attached Figure Description

[0022] Figure 1 This is a flowchart of the collaborative adaptive posterior enhancement long-tail recommendation method in an embodiment of the present invention; Figure 2 This is a schematic diagram of the collaborative adaptive posterior enhancement long-tail recommendation method in an embodiment of the present invention; Figure 3 This is a structural block diagram of the collaborative adaptive posterior augmentation long-tail recommendation system in an embodiment of the present invention; Figure 4 This is a structural block diagram of the electronic device in an embodiment of the present invention. Detailed Implementation

[0023] To better understand the purpose, technical solution, and advantages of this application, the application is described and explained below in conjunction with the accompanying drawings and embodiments.

[0024] Unless otherwise defined, the technical or scientific terms used in this application shall have the general meaning understood by one of ordinary skill in the art to which this application pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this application do not indicate quantitative limitation and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this application are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. The terms "first," "second," "third," etc., used in this application are merely to distinguish similar objects and do not represent a specific order of objects.

[0025] To facilitate understanding, the present invention will first be briefly described as follows: like Figure 1 As shown, Figure 1 The schematic diagram of this invention is shown. After the user's historical interaction sequence is searched by the constraint bundle of the backbone generation model, a candidate item list is generated. The CAPE module calculates the collaborative prior for long-tail items in the candidate list and completes the log-space posterior correction based on the offline pre-computed item co-occurrence matrix. The adaptive gating module controls the correction intensity according to the current user's historical length. The ranking condition intervention module detects the Top-1 score gap and rearranges the rankings from 2 to K while protecting the Top-1 prediction, and finally outputs a recommendation list with improved long-tail coverage.

[0026] Specifically, such as Figure 1 and Figure 2 As shown in the figure, a collaborative adaptive posterior augmentation long-tail recommendation method provided in this embodiment of the invention includes the following steps: S1. Obtain the list of candidate items generated for the user by the generative recommendation model through constraint bundle search and the log probability score of each candidate item; It should be noted that, firstly, let the set of items be I, and the user... The historical interaction sequence is , Given the historical length, the generative recommender defines the probability distribution on items by constraining semantic ID decoding. Based on the training frequency, the bottom 80% of items by training frequency are defined as the long-tail item set (Tail), and the rest are defined as the head item set (Head).

[0027] Bundle-induced long-tail collapse exhibits a two-layer failure structure: the first layer is an exclusion layer, where, at bundle width B, the probability of long-tail target items appearing in the candidate set (bundle recall) is significantly lower than that of head items, and this gap persists across all bundle widths, making it impossible to completely eliminate by increasing the bundle width; the second layer is a score suppression layer, where even if long-tail items enter the candidate set, their ranking is systematically suppressed due to the cumulative popularity bias during multi-step decoding. Experiments with a bundle width B=50 validated this, showing that 70.0% (Industrial dataset) and 41.4% (Office dataset) of long-tail items entering the candidate set ranked below 10th, indicating that score suppression is the main manageable bottleneck at the current operational point.

[0028] S2. Based on the user's historical interaction sequence, the collaborative prior of each candidate item is calculated using the pre-computed item co-occurrence matrix, and the corrected score of the long-tail candidate item is obtained through log space posterior correction. For example, suppose Let be the sparse item co-occurrence matrix computed from the training interactions, where Statistics on both item i and item i The number of users who interacted with any candidate item. Laplace smoothing computation of user-level collaboration priors: , ; in, Candidate items In users Normalized cooperative prior probabilities under the following conditions; For candidate items The average co-occurrence at the user level; The normalization constant is the mean co-occurrence value of all candidate items. Summation, making This constitutes a valid probability distribution.

[0029] To ensure that the correction only applies to long-tail candidate items, a normalized long-tail logarithm prior is defined: ; in The smallest non-zero prior value among all candidates ensures ≥0 for all items Established, and guaranteed =0 holds true for all head items, thus ensuring that the correction signal is specific to long-tail candidates. The final corrected rearrangement score is: ; in, Represents the logarithmic probability of the model. This is a normalized long-tail logarithm prior that applies only to long-tail candidate items. To correct the intensity hyperparameter, the collaborative prior is added to the model's log-likelihood in log space, which is equivalent to normalization of the Bayesian posterior. This can theoretically reduce the KL divergence between the corrected distribution and the true user preference distribution.

[0030] S3. Based on the user's historical length, dynamically adjust the correction strength of the correction score using an adaptive gating function; Specifically, the variance of the collaborative prior estimate is inversely proportional to the length of the user history; the shorter the history, the less reliable the prior estimate. To address this, a saturated adaptive gating function is introduced: in, To globally correct the intensity hyperparameter; For users after gating adjustment The actual correction strength; For linearly truncated gating functions, the correction strength is correlated with the history length; The preset saturation threshold, For the length of user history, when The gating saturation is set to 1 to maintain the full correction strength. This design makes the correction strength positively correlated with the reliability of the collaborative prior, effectively preventing cold-start users from being subjected to excessive interference due to unreliable priors.

[0031] S4. Apply the ranking condition intervention strategy, keep the original Top-1 prediction of the locked bundle search unchanged, and only rearrange the candidate items ranked from 2nd to Kth according to the modified score, and output the final recommendation list.

[0032] The ranking criteria intervention strategy specifically includes: Compute the difference in log probability scores between the original Top-1 and Top-2 candidates using beam search. Delta 12 (u) ; Calculate the maximum correction magnitude λ·h among all long-tail candidate items. max ; If λ·h max ≥Δ 12 If (u), then the final ranking of the original Top-1 candidate is locked as first place, and the corrected score is applied to the candidate items ranked 2nd to Kth in the original ranking for re-ranking. Otherwise, apply the corrected scores to all candidate items and rearrange them.

[0033] For example, let Δ12(u) be the difference in log probability scores between the Top-1 and Top-2 candidates in the beam search, h max The maximum value of the long-tailed logarithmic prior among all candidates, when the correction magnitude λ·h max Exceed Delta 12 (u) At this time, long-tail items may surpass the original Top-1 prediction. Therefore, a ranking condition intervention strategy locks the Top-1 prediction as the optimal candidate for the original search, and only applies corrected scores to the positions ranked 2 to K for re-ranking. Theoretical analysis shows that this strategy can constrain the HR@1 loss within a controllable range, and the upper bound of the change in HR@K (K≥2) is only related to the proportion of long-tail items in the test set.

[0034] In summary, this method includes the following stages: Offline construction phase: Collect user historical interaction sequences, map item IDs to semantic ID sequences, extract the co-occurrence frequency of all item pairs, and construct a sparse co-occurrence matrix.

[0035] Parameter configuration stage: Set the correction intensity λ and the gate saturation threshold. and ranking protection strategies.

[0036] Online inference phase: The backbone generation model uses the original constraint bundle search to generate a list of candidate items and corresponding probability scores.

[0037] Posterior correction phase: For a given user query co-occurrence matrix, calculate the collaborative prior of each long-tail candidate, apply gating coefficients, and generate the corrected rearranged score.

[0038] Ranking intervention phase: Detect the score difference between Top-1 and Top-2. If the correction exceeds the difference, lock Top-1 and only apply the corrected score to the rankings from 2 to K to re-rank them, and output the final recommendation list.

[0039] This method has the following advantages and positive effects: (1) Training-independent and plug-and-play. This invention does not modify any parameters of the backbone generation model, does not require additional training data or optimization process, and can be directly integrated into the existing generative recommendation inference process as a post-processing plug-in. It does not require structural modification of the existing system and has extremely low deployment cost.

[0040] (2) Fundamentally solve long-tail bias from the inference stage. Existing debiasing methods mainly target the training stage and cannot prevent bundle search from reintroducing and accumulating popularity bias during inference. This invention directly intervenes in candidate scores during inference, using a Bayesian posterior correction framework to reduce the bias between the model distribution and the real user preference distribution, thereby eliminating the cascading amplification effect of bundle search from a mechanistic perspective.

[0041] (3) The correction signal is precisely targeted at long-tail candidates. The long-tail indicator function ensures that the collaborative prior correction signal is applied only to long-tail candidate items, while the scores of head items remain unchanged, thus avoiding interference with the prediction of high-confidence popular items.

[0042] (4) Adaptive protection for cold starts and sparse users. The inverse variance weighted gating mechanism automatically adjusts the correction strength according to the length of the user's history. Users with short interaction history are protected and will not generate excessive error correction due to unreliable prior estimates of collaboration, thus improving the robustness of the system.

[0043] (5) Strictly maintain the overall Top-1 recommendation accuracy. The ranking condition intervention strategy ensures that HR@1 is not reduced due to long-tail correction. Compared with competitive methods that need to sacrifice overall accuracy for long-tail coverage, it has a significant advantage and is suitable for online recommendation scenarios with strict requirements for Top-1 accuracy.

[0044] (6) The experimental results are significant. On two Amazon standard recommendation benchmark datasets, the long-tail item hit rate of this invention is up to 72.1% higher than that of the original bundle search, and the overall recommendation hit rate is also improved accordingly, while HR@1 is maintained at the baseline level. The specific results are shown in Table 1.

[0045] Table 1: Recommended metrics for different data domains and baselines (MiniOneRec backbone model, bundle width B=50)

[0046] Example 1: An e-commerce platform encodes its product catalog as a multi-token semantic ID sequence and uses a large language model as a generative recommender. When a user requests a recommendation, the backbone model generates a candidate product list through constraint-based search. The CAPE module, based on an offline pre-computed item co-occurrence matrix, completes the posterior score correction of the candidate list within milliseconds, significantly improving the exposure opportunities of niche products while maintaining the recall accuracy of popular products, effectively improving the fairness of long-tail product distribution on the platform.

[0047] Example 2: A content recommendation platform has a large number of new users with short historical interaction sequences. CAPE's adaptive gating mechanism automatically detects that the historical length of these users is insufficient to meet the threshold, and proportionally compresses the correction intensity to avoid introducing noise due to unreliable collaborative priors. As users continue to interact and their history gradually accumulates, the correction intensity increases linearly, enabling the system to provide stable recommendation services throughout the user's entire lifecycle.

[0048] Example 3: A large amount of content on a short video platform belongs to long-tail videos. The generative recommendation system suffers from severe underexposure of long-tail content because popular videos dominate the training data. After deploying this invention, the system provides collaborative prior enhancements for long-tail videos based on item co-occurrence statistics of users' historical viewing sequences. Combined with ranking condition intervention strategies, this ensures that the Top-1 recommendation of popular videos remains unaffected, improving the distribution coverage of long-tail content while maintaining the fulfillment of users' primary needs for popular content.

[0049] In summary, this method constructs a collaborative prior based on the item co-occurrence statistics in user history interactions after generating a candidate list through beam search. This prior is then fused with the model's beam search score in logarithmic space to achieve posterior enhancement of long-tail item scores. An adaptive gating mechanism controls the correction intensity based on the length of user history. Finally, a ranking condition intervention strategy protects Top-1 predictions, concentrating corrections on rankings 2 to K. This achieves a systematic improvement in long-tail item recommendations without introducing any training overhead.

Claims

1. A collaborative adaptive posterior augmentation long-tail recommendation method, characterized in that, Includes the following steps: Obtain the list of candidate items generated by the generative recommendation model for the user through constraint bundle search, and the log probability score of each candidate item; Based on the user's historical interaction sequence, the collaborative prior of each candidate item is calculated using a pre-computed item co-occurrence matrix, and the corrected score of the long-tail candidate item is obtained through log space posterior correction. The correction strength of the correction score is dynamically adjusted based on the user's historical length using an adaptive gating function; By applying a ranking condition intervention strategy, the original Top-1 prediction of the locked bundle search remains unchanged, and only the candidate items ranked from 2nd to Kth are re-ranked according to the modified scores to output the final recommendation list; The corrected score is expressed as: ; in, Represents the logarithmic probability of the model. This is a normalized long-tail logarithm prior that applies only to long-tail candidate items. To correct the intensity hyperparameter.

2. The collaborative adaptive posterior augmentation long-tail recommendation method according to claim 1, characterized in that, The method for constructing the collaborative prior includes: Construct a sparse item co-occurrence matrix from the training interaction data. ; For users Based on its historical interaction sequence Laplace smoothing was used to calculate the user-level co-occurrence mean. ,in Statistics on items and items The number of users who generated interaction, The length of the user's historical interaction sequence; Normalize the mean of user-level co-occurrence to obtain the collaborative prior distribution: ， ； in, Candidate items In users Normalized cooperative prior probabilities under the following conditions; For candidate items The average co-occurrence at the user level; The normalization constant is the mean co-occurrence value of all candidate items. Summation, making This constitutes a valid probability distribution.

3. The collaborative adaptive posterior augmentation long-tail recommendation method according to claim 1, characterized in that, The normalized long-tailed logarithmic prior Defined as: ； in, For long-tail indicator functions, Let the smallest non-zero collaborative prior value among all candidate items be such that... And head items .

4. The collaborative adaptive posterior augmentation long-tail recommendation method according to claim 1, characterized in that, The adaptive gating function is a saturated linear gating function: ； in, To globally correct the intensity hyperparameter; For users after gating adjustment The actual correction strength; For linearly truncated gating functions, the correction strength is correlated with the history length; The preset saturation threshold, For the length of user history, when When the gating saturation is 1, the complete correction strength is maintained.

5. The collaborative adaptive posterior augmentation long-tail recommendation method according to claim 1, characterized in that, The ranking criteria intervention strategy specifically includes: Compute the difference Δ between the log probability scores of the original Top-1 and Top-2 candidates in the beam search. 12 (u); Calculate the maximum correction magnitude λ·h among all long-tail candidate items. max ; If λ·h max ≥Δ 12 If (u), then the final ranking of the original Top-1 candidate is locked as first place, and the corrected score is applied to the candidate items ranked 2nd to Kth in the original ranking for re-ranking. Otherwise, apply the corrected scores to all candidate items and rearrange them.

6. The collaborative adaptive posterior augmentation long-tail recommendation method according to claim 1, characterized in that, The method does not modify any parameters of the generative recommendation model, does not rely on additional training data or optimization processes, and can be integrated into the inference process of existing generative recommendation systems as a post-processing plugin.

7. The collaborative adaptive posterior augmentation long-tail recommendation method according to claim 1, characterized in that, It also includes an offline construction phase: collecting user interaction sequences, mapping items to semantic ID sequences, counting the frequency of occurrence of all item pairs, and constructing and storing a sparse item co-occurrence matrix.

8. The collaborative adaptive posterior augmentation long-tail recommendation method according to claim 1, characterized in that, The generative recommendation model is an autoregressive generative model based on semantic IDs, which generates a sequence of semantic IDs for candidate items and their corresponding cumulative log probabilities by decoding through constraint bundle search.

9. The collaborative adaptive posterior augmentation long-tail recommendation method according to claim 1, characterized in that, The long-tail items are defined as items whose interaction frequency ranks in the bottom 80% of the training data, and the head items are items in the top 20% of the data.

10. A collaborative adaptive posterior augmentation long-tail recommendation system, characterized in that, include: The co-occurrence matrix storage module is used to store the sparse item co-occurrence matrix obtained in advance from the training interaction data; The collaborative prior calculation module is used to query the co-occurrence matrix based on the user's historical interaction sequence, calculate the user-level co-occurrence mean and normalized collaborative prior of each candidate item, and generate a normalized long-tail logarithmic prior that only applies to long-tail candidate items. The adaptive gating module is used to calculate the saturated linear gating coefficient based on the user's historical data and dynamically adjust the correction strength. The posterior correction module is used to add the logarithmic probability scores of candidate items output by the model to the gated long-tailed logarithmic prior in the logarithmic space to obtain the corrected rearrangement scores. The ranking condition intervention module is used to detect the difference between the original Top-1 and Top-2 scores in the beam search. When the correction magnitude may change the Top-1 prediction, the original Top-1 prediction is locked, and the corrected scores are applied only to the candidate items ranked 2nd to Kth to re-rank them, and the final recommendation list is output.

Images