### Personalized recommendation scenarios FAQ :::tip \*说明 This document is a troubleshooting guide for the recommendation system, designed to help users quickly locate the cause and resolve common issues, such as empty recommendation results and poor personalization performance. This article helps you understand troubleshooting methods and specific procedures for common problems, thereby improving the effectiveness and stability of the recommendation system. To experience more Viking AI Search capabilities, go to: [Viking AI Search, an all-in-one intelligent engine for enterprise search, recommendation, and Q&A](https://www.volcengine.com/product/AI-Search-Rec) ::: ### Empty recommendation results For details, see the help document: [Why does the recommendation return no items?](https://www.volcengine.com/docs/85296/2124616?lang=zh#bd9b3c9f) You can also follow the process below to troubleshoot. |Troubleshooting module |Troubleshooting logic and recommendations | |---|---| |User event data quality|1. **Item ID**: Can be found in the item table|\ | |2. **User event time**: Within 7 days (events older than 7 days can also be sent, but events within the last 7 days are required).|\ | |3. **Event type**: Both exposure and non\-exposure positive user events (such as Click, Favorite, and Purchase) are required.|\ | |4. **Number of user events**: As many as possible, with a minimum of 1,000. | |Procedure |Procedure: First, on the Applications page in the console, associate at least one item dataset with your application and Click Apply. Then, on the same page, associate a User event dataset and Click Apply. Finally, Click the Effective configuration button at the top of the page to apply all changes.|\ | |Note: The User event dataset must be associated and applied after the Image and text dataset is applied. You must follow this order because User event data (such as a Click) needs to be associated with specific Items. If the item dataset is not yet effective, the system will not be able to parse the Item information in the User event data, causing the association to fail. | |Define recommendable items|

|\ | |

|\ | ||\ | |The main issue is that no Items meet all of the above conditions. Let's break down the possible reasons:|\ | ||\ | |* The OR logic was incorrectly configured as AND.|\ | |* A Field in the Item table has a transmission problem, and the actual Value does not match the defined format.|\ | |* No Item meets all the conditions.|\ | ||\ | | | |Remove duplicate recommendations|Filtered out by exposure and delivery deduplication|\ | |

|\ | |

|\ | ||\ | | | |Fallback strategy |Reasonableness of the fallback strategy|\ | |

|\ | |

|\ | ||\ | | | |Item data quality |1. **The item's tag field** must have a value. The field designated for tags is inferred by the Volcano Engine model. If you have a more suitable field, you can contact Volcano Engine to request a change.|\ | ||\ | |

|\ | |

|\ | ||\ | | | ### Slow recommendations Recommendations may be slow if a timeout occurs when requesting item attribute information during item retrieval. This can happen if an attribute value is very large. Check the item table for fields with very large values, such as the following: |Check item |Investigation logic and recommendation | |---|---| |Multi\-modal reranking|Is the number of items for multi\-modal reranking too large? A configuration of 20 is recommended.|\ | |

|\ | |

|\ | ||\ | | | |Are there large fields used for document search? |

|\ | |

|\ | | | ### Empty user profile in recommendation scenarios > Similar issue: test user interest tags are empty The calculation logic for user profile tags is as follows: It counts the item IDs from a user's positive behavior within the last 7 days and uses them to query the corresponding user interest tags field in the item table. The user interest tags field is a field intelligently selected by the Huoshan large model after analyzing the item dataset. **Suggestion:** First, check if the user has any positive behavior within the last 7 days. If so, check if the item from the positive behavior can be found in the item table. If the item can also be found, contact Huoshan technical support to check if the intelligently selected item field is appropriate. If not, adjust it manually. ### Recommendation results are not personalized enough > Description: > The expected recommendation behavior is: if a user clicks or plays an item of a certain type, they should later be recommended items of the same type. > If the recommendation content does not meet expectations, there may be an issue with the configuration. Currently, three item retrieval paths return personalized items, as shown in the table below. To achieve strong personalized recommendation results, you need to optimize the data quality according to the logic of each item retrieval path. **In particular, optimize user interest item retrieval and large model item retrieval**. |Item retrieval channel |Item retrieval logic |Customer\-side confirmed optimization | |---|---|---| |User interest item retrieval|1. Calculate trending items under each tag based on all user event data.|1. User event data|\ | |2. Get the list of tags for items associated with a user's positive behavior in the last 7 days, then retrieve trending items for each tag in the list.| 1. Large data volume|\ | || 2. Includes Exposure and Positive behavior (e.g., Click, Purchase, play, etc.)|\ | | | 3. Items must be findable in the item data|\ | |**Note: The tags here are from a field that the platform's large model automatically selects after analyzing the item table fields. Customers can infer which field the tags come from by viewing the tag values displayed below.** |2. Item data|\ | |

| 1. Criteria for evaluating the interest dimension field automatically selected by Volcano Engine|\ | |

| 1. The field value is consistent with the business's trade\-offs regarding user interest dimensions. For example, commonly selected fields include product category and information tags.|\ | || 2. The field has high coverage, meaning that most items have a value for this field.|\ | | | 3. The field's enumerated values should be distinct (avoid having most or even all items share a single tag value) and not too sparse (avoid having a large number of values where most correspond to very few items).|\ | | | 2. If there is a more suitable tag field, contact Volcano Engine technical support to make an adjustment.|\ | | ||\ | | | |\ | | | |\ | | | | |Large model item retrieval|1. Creates an embedding for each item based on the information in the fields selected as Searchable (Text)|1. Item data fields for Document search|\ | |2. Retrieves a list of items with which the user has had positive behavior in the last 7 days, creates embeddings for these items, and then performs a vector search for similar items | 1. Select fields that contain core information, such as Title and categories|\ | | | 2. Do not select fields with non\-core information to prevent miscellaneous data from diluting the core information and causing irrelevant items to be retrieved | |ItemCF item retrieval|1. Calculates item\-to\-item similarity based on the full user event data|1. User event data|\ | |2. Retrieves a list of items from the user's positive behaviors over the last 7 days. For each item on the list, it then retrieves other highly similar items.| 1. Large data volume|\ | || 2. Includes Exposure and positive behavior (e.g., Click, Purchase, Play, etc.) |\ | | | | ### Similar items for parent items are not available on detail pages |Checklist item |Troubleshooting logic | |---|---| |API parameter passing |Check: whether the parameter format is correct. Correct parameter passing is as follows:|\ | |

|\ | |

|\ | ||\ | | | |Parent item |Check: the parent item ID must be within the item dataset. | ### The recommendation returns duplicate items |Troubleshooting module |Troubleshooting logic | |---|---| |Remove duplicate recommendations |If the deduplication configuration is too small, only a few recently recommended items are filtered out with each new recommendation, which can cause previously recommended items to be recommended again.|\ | |(The maximum time window for exposure deduplication is 7 days)|\ | |

|\ | |

|\ | ||\ | | | |User event data timeliness |Check whether exposure user event data is synchronized to the Volcengine platform in **real time**.|\ | |If latency is high, Exposed Item Deduplication cannot retrieve exposed items, causing the process to fail (Suggestion: Improve the timeliness of user event data and add Delivered Item Deduplication). | ### Diversification is not taking effect |Troubleshooting items |Troubleshooting logic and recommendations | |---|---| |**Confirm that it has indeed not taken effect.** |1. Place the diversification field in output_fields and use the information returned by the API to verify if diversification is not applied.|\ | |2. On the product interface, it may seem that diversification is not working even when it is effective. This is mainly caused by the cross\-batch logic. The detailed logic is as follows:|\ | | 1. The current effective scope for diversification is Within each batch. For example, if one batch returns 5 items, and the configured diversification logic is|\ | ||\ | |

|\ | |

|\ | ||\ | ||\ | | 2. then Within each batch, the condition of displaying at most 1 item of the same type for every 4 items can be met. However, the last Item of the first batch and the first Item of the second batch might be of the same dimension. This leads to a concentration of similar items when browsing the application, which is an expected behavior according to the functional logic.|\ | ||\ | |

|\ | |

|\ | ||\ | | | |**Is the diversification level too high**|If the content is too diversified, for example, by not allowing the same dimension to appear over a long consecutive sequence, diversification may fail because the items from Item retrieval do not have enough variety.|\ | |Furthermore, in personalized Recommendation practice, each user has specific preferences for certain dimensions. For example, a user might recently want to buy a certain style of down jacket, or another user might like the works of a specific author. In such cases, the Recommendation system will Return more items of this type to the user.|\ | |Finally, because diversification is only effective Within a single batch, if one batch returns 10 items and diversification is configured to Display at most N items within any 20 consecutive items, it will also likely lead to unexpected results.|\ | |Suggestion:|\ | |

|\ | |

|\ | ||\ | | | |Is the diversification Field appropriate |1. Nested fields cannot be used for unwinding.|\ | |2. If a field has a null value or no unwind field, it is treated as a single dimension. Therefore, if the unwind field contains many null values, it increases the probability of unwinding failure. | ### The recommendation filter is not taking effect A dynamic filter requires both platform configuration and parameters passed via the API to take effect. If either is missing, the filter will not work, and the missing component will be indicated in the `omitted_params` field of the API response.

You can use the table below to investigate the possible reasons in order: |Cause classification |Troubleshooting logic and recommendations | |---|---| |Passed in the API call, but not configured in the filter policy.|![图片](https://p9-arcosite.byteimg.com/obj/tos-cn-i-goo7wpa0wc/590f5de277e94921abf3790c368871f8 =395x) |\ | |![图片](https://p9-arcosite.byteimg.com/obj/tos-cn-i-goo7wpa0wc/79e2d5c87ad94ef68a723e501dbdeeb5 =846x) |\ | | |\ | | | |A dynamic filter is configured in the filter policy, but not passed in the API parameters. |![图片](https://p9-arcosite.byteimg.com/obj/tos-cn-i-goo7wpa0wc/7e85343fb22b43b698519c863ce9571d =1300x) |\ | | | |Invalid dynamic filter parameters |

|\ | |

|\ | ||\ | |

|\ | |

|\ | ||\ | | | |Logical condition check |In the filter expression logic, if `not found in` is followed by a list, the condition is met only if it applies to all elements, not just if an intersection exists.|\ | |If you want to filter based on matching any single element, use multiple expressions.|\ | | | |Others|Check if the configured logic is OR|\ |Incorrect logical condition |

|\ | |

|\ | ||\ | | | :::tip \*说明 To learn more about the capabilities of Viking AI Search, see [Viking AI Search, the integrated intelligent engine for enterprise Search, Recommendation, and Q&A](https://www.volcengine.com/product/AI-Search-Rec). :::