Artificial Intelligence Recommendation:Sorting configuration

How to configure

The sorting configuration corresponds to `SortConfs` in the configuration overview. `SortConfs` is a `[]object` structure that lets you configure multiple sorting policies. PAI-Rec provides the following built-in policies: BoostScoreSort, BoostScoreByWeight, ItemRankScore, DiversityRuleSort, DPPSort, and MultiRecallMixSort.

Common sorting configurations

Each sorting configuration uses a subset of the common configurations. This section explains these configurations to avoid repetition.

Configuration example:

{
    "SortConfs": [
        {
            "Name": "",
            "SortType": ""
        }
    ]
}

Boost score sort (BoostScoreSort)

After the fine-grained sorting model is invoked, each item receives a score. Based on business requirements, you may need to modify this score by boosting or demoting it.

The boost or demotion operation consists of two parts:

• Set conditional rules. You can use item or user properties, such as category or gender, to determine whether the conditions are met.

• Set the boost or demotion expression. Currently, you can only set expressions for the score, such as `score × 1.2` or `score × 0.5`.

Configuration example:

{
    "SortConfs": [
        {
            "Name": "BoostScoreSort",
            "SortType": "BoostScoreSort",
            "Debug": false,
            "BoostScoreConditions": [
                {
                    "Conditions": [
                        {
                            "Name": "sex",
                            "Domain": "item",
                            "Type": "string",
                            "Value": "gender",
                            "Operator": "equal"
                        }
                    ],
                    "Expression": "score * 2"
                }
            ]
        }
    ]
}

This configuration multiplies the score by 2 for items where the `sex` feature is `male`.

The `FilterParamConfig` configuration is as follows:

For more information about conditional settings, see Adjust count filter (AdjustCountFilter).

Boost score by weight (BoostScoreByWeight)

When you boost or demote item scores, different items can have different weights. The weight is a field in the item table and is used to adjust the score.

Score formula: `weight × item.score`

Configuration example:

{
    "SortConfs": [
        {
            "Name": "BoostScoreByWeight",
            "SortType": "BoostScoreByWeight",
            "TimeInterval": 172800,
            "BoostScoreByWeightDao": {
                "AdapterType": "hologres",
                "HologresName": "pai_rec",
                "HologresTableName": "test",
                "ItemFieldName": "item_id",
                "WeightFieldName": "weight"
            }
        }
    ]
}

BoostScoreByWeightDao

Item rank score (ItemRankScore)

`ItemRankScore` sorts items in descending order based on their scores. This feature is built into the DPI engine and can be used directly in `SortNames`.

Diversity rule sort (DiversityRuleSort)

When outputting recommendation results, consider not only user interests but also item diversity. This lets you mix items from different categories and with different properties in the output.

You can configure the rules as follows:

Recommendation diversity rules

Glossary:

1. Diversification dimension: The item property used for diversification, such as category, author, or tag.

2. Diversification policy:

   • Minimum interval k: An item from a specific diversification dimension can appear consecutively at most k times.

   • Maximum frequency m: Within a window of size n, an item from the same diversification dimension cannot appear more than m times.

Diversification logic:

• Diversification rules apply only to the results of a single request. Cross-request diversity is not considered.

• You can configure multiple diversification dimensions.

• For each diversification dimension, you can configure multiple diversification policies, each with different parameters (k, m, n).

Configuration example:

{
    "SortConfs": [
        {
            "Name": "DiversityRuleSort",
            "SortType": "DiversityRuleSort",
            "DiversitySize": 100,
            "DiversityRules": [
                {
                    "Dimensions": ["spfl"],
                    "WindowSize": 10,
                    "FrequencySize": 1
                }
            ],
            "ExcludeRecalls": [
                "ColdStartVideoVectorRecall",
                "LinUcbRecall_default2"
            ],
            "Conditions": [
                {
                    "Name": "spflPick",
                    "Domain": "user",
                    "Type": "string",
                    "Value": "",
                    "Operator": "equal"
                }
            ]
        }
    ]
}

This sort must be used with the `ExcludeRecalls` parameter.

DiversityRules

See the following example for exclusion rules and search depth.

Items that meet the condition `tag=t1` do not appear in output positions 1, 2, 3, or 4. The items that appear in positions 1, 2, 3, and 4 must still comply with the diversification rules, but they cannot have `tag=t1`.

{
    "Name": "DiversityRuleSort",
    "SortType": "DiversityRuleSort",
    "DiversityRules": [
        {
            "Dimensions": [
                "tag"
            ],
            "WindowSize": 5,
            "FrequencySize": 1
        }
    ],
    "ExclusionRules": [
        {
            "Positions": [
                1,2,3,4
            ],
            "Conditions": [
                {
                    "Name": "tag",
                    "Domain": "item",
                    "Type": "string",
                    "Value": "t1",
                    "Operator": "equal"
                }
            ]
        }
    ],
    "ExploreItemSize" : 200
}

When searching for items that satisfy diversity rules, an item is output by default only if it meets all rules. If any diversity rule is not met, the system searches for the next item until one that meets all rules is found. If no item in the candidate set meets all rules, the first item searched is selected for output.

Another policy is available. You can add weights to the diversity rules. If no item in the candidate set meets all rules, the item that satisfies the rules with the highest total weight is selected for output. If multiple items have the same highest total weight, the one that appears earliest in the list is chosen.

See the following example for a configuration with weights:

{
    "Name": "DiversityRuleSort",
    "SortType": "DiversityRuleSort",
    "DiversityRules": [
        {
            "Dimensions": [
                "tag"
            ],
            "WindowSize": 5,
            "FrequencySize": 1,
            "Weight": 1
        },
        {
            "Dimensions": [
                "category"
            ],
            "WindowSize": 3,
            "FrequencySize": 1,
             "Weight": 3
        }
    ],
    "ExclusionRules": [
        {
            "Positions": [
                1
            ],
            "Conditions": [
                {
                    "Name": "tag",
                    "Domain": "item",
                    "Type": "string",
                    "Value": "t1",
                    "Operator": "equal"
                }
            ]
        }
    ]
}

DPPSort

For more information about the DPP (Determinantal Point Process) diversity algorithm, see An Intuitive Understanding of the Determinantal Point Process-based Algorithm for Improving Recommendation Diversity.

Prerequisites: To use the DPP algorithm, you must have item embedding vectors. These vectors must represent the content of the items. The similarity of the embeddings must represent similarity at the item content level, not at other levels such as behavior. For example:

• Recommended: Item image embeddings, text description embeddings, or embeddings generated from a combination of static item content such as categories and properties.

• Not recommended: Embeddings trained from models based on user behavioral data.

Essentially, the dimensions you want to diversify must be reflected in the embeddings. For example, if you want the recommendation list to have diversity in the product price dimension, the price feature must be included when training the model to obtain the embedding vectors. Otherwise, you will not achieve the desired effect.

Configuration example:

{
    "SortConfs": [
        {
            "Name": "DPPSort",
            "SortType": "DPPSort",
            "DPPConf": {
                "Name": "DPPSort",
                "DaoConf": {
                    "AdapterType": "hologres",
                    "HologresName": "geeko_rec"
                },
                "TableName": "item_embedding_metric_learning",
                "TableSuffixParam": "embedding_date",
                "TablePKey": "product_id",
                "EmbeddingColumn": "embedding",
                "Alpha": 4.5,
                "NormalizeEmb": "false",
                "WindowSize": 10
            }
        }
    ]
}

DPPConf

SSDSort

For more information about the SSD (Structured Self-Distillation) diversity algorithm, see Improving the Diversity of Recommendation Results: An Analysis of MMR/DPP/SSD Principles.

Prerequisites: To use the SSD algorithm, you must have item embedding vectors. These vectors must represent the content of the items. The similarity of the embeddings must represent similarity at the item content level, not at other levels such as behavior. For example:

Recommended: Item image embeddings, text description embeddings, or embeddings generated from a combination of static item content such as categories and properties.

Not recommended: Embeddings trained from models based on user behavioral data.

Essentially, the dimensions you want to diversify must be reflected in the embeddings. For example, if you want the recommendation list to have diversity in the product price dimension, the price feature must be included when training the model to obtain the embedding vectors. Otherwise, you will not achieve the desired effect.

Configuration example:

{
    "SortConfs": [
        {
            "Name": "SSDSort",
            "SortType": "SSDSort",
            "SSDConf": {
                "Name": "SSDSort",
                "DaoConf": {
                    "AdapterType": "hologres",
                    "HologresName": "geeko_rec"
                },
                "TableName": "item_embedding_metric_learning",
                "TablePKey": "item_id",
                "EmbeddingColumn": "embedding",
                "Gamma": 0.25,
                "UseSSDStar": true,
                "NormalizeEmb": "false",
                "MinScorePercent": 0.1,
                "CandidateCount": 200,
                "WindowSize": 5
            }
        }
    ]
}

SSDConf

Multi-channel recall sort (MultiRecallMixSort)

Typically, you will have many recall channels. Sometimes, based on business requirements, you need to mix the output based on the recall type. For example:

• There are exposure quantity requirements for cold-start recall.

• There are position requirements for a specific recall channel.

This may also include multiple mixing rules.

The configuration is as follows:

{
    "SortConfs": [
        {
            "Name": "MixSort",
            "SortType": "MultiRecallMixSort",
            "RemainItem": false,
            "MixSortRules": [
                {
                    "MixStrategy": "random_position",
                    "NumberRate": 0.1,
                    "RecallNames": [
                        "OTSGlobalHot"
                    ]
                },
                {
                    "MixStrategy": "fix_position",
                    "Positions": [1,3,5],
                    "RecallNames": [
                        "RecallName1"
                    ]
                }
            ]
        }
    ]
}

In addition to using recall names to match and filter items, you can also use conditional filtering to select items for exposure.

{
    "SortConfs": [
        {
            "Name": "MixSortByItemFeature",
            "SortType": "MultiRecallMixSort",
            "RemainItem": false,
            "MixSortRules": [
                {
                    "MixStrategy": "random_position",
                    "NumberRate": 0.1,
                    "Conditions": [
                        {
                            "Name": "gender",
                            "Domain": "item",
                            "Type": "string",
                            "Value": "man",
                            "Operator": "equal"
                        }
                    ]
                }
            ]
        }
    ]
}

How to use

The sorting configuration is similar to the recall configuration. After configuration, you provide a `SortNames` parameter for use in different scenarios. `SortNames` is a `Map[string]object` structure where the key is the scenario, and each scenario corresponds to a set of sorting policies.

{
    "SortNames": {
        "${scene_name}": [
            "ItemRankScore"
        ]
    }
}

• `${scene_name}` is the scenario name. If you want to use the same configuration for multiple scenarios, you can use `default`.

• `ItemRankScore`: This parameter is the custom name of the sort defined in `SortConfs`.