Unlearning documentation

[kmaehashi]

We now have unlearning feature for:

• Classifier
• Recommender (nearest_neighbor_recommender)
• Anomaly (light_lof)

We should cover the following topics:

• List of algorithms that can be used for unlearning
• Behavior for each algorithm

Preferably with example config.

Note that we changed recommender's config syntax when using nearest_neighbor_recommender.

[kmaehashi]

Finally the config looks like this.

Classifier:

{
  "converter" : {
    "string_filter_types" : {},
    "string_filter_rules" : [],
    "num_filter_types" : {},
    "num_filter_rules" : [],
    "string_types" : {},
    "string_rules" : [
      { "key" : "*", "type" : "str", "sample_weight" : "bin", "global_weight" : "bin" }
    ],
    "num_types" : {},
    "num_rules" : [
      { "key" : "*", "type" : "num" }
    ]
  },
  "method" : "PA1",
  "parameter": {
    "regularization_weight" : 1.0,
    "unlearner": "lru",
    "unlearner_parameter": {
      "max_size": 10
    }
  }
}

Recommender:

{
  "converter" : {
    "string_filter_types": {},
    "string_filter_rules":[],
    "num_filter_types": {},
    "num_filter_rules": [],
    "string_types": {},
    "string_rules":[
      {"key" : "*", "type" : "str", "sample_weight":"bin", "global_weight" : "bin"}
    ],
    "num_types": {},
    "num_rules": [
      {"key" : "*", "type" : "num"}
    ]
  },
  "parameter" : {
    "parameter": {
     "hash_num" : 64
    },
    "method": "lsh",
    "unlearner": "lru",
    "unlearner_parameter": {
       "max_size": 10
    }
  },
  "method": "nearest_neighbor_recommender"
}

Anomaly:

{
  "converter" : {
    "string_filter_types" : {},
    "string_filter_rules" : [],
    "num_filter_types" : {},
    "num_filter_rules" : [],
    "string_types" : {},
    "string_rules" : [
      { "key" : "*", "type" : "str", "sample_weight" : "bin", "global_weight" : "bin" }
    ],
    "num_types" : {},
    "num_rules" : [
      { "key" : "*", "type" : "num" }
    ]
  },
  "parameter" : {
    "nearest_neighbor_num" : 10,
    "reverse_nearest_neighbor_num" : 30,
    "method" : "euclid_lsh",
    "parameter" : {
      "hash_num" : 64
    },
    "unlearner": "lru",
    "unlearner_parameter": {
      "max_size": 10
    }
  },
  "method" : "light_lof"
}

[kmaehashi]

We should also write about delete_label API added in classifier.

[kmaehashi]

Fixed via #144