Prediction API

Dynamic Imputing of Missing Value for Predictions

Real world data often times contains missing values. While these missing values are handled gracefully by some machine learning algorithms, others, require all values to be present in order to generate a prediction. For the latter SymetryML can impute the missing values in one of the following ways:

• Replace the missing value by the average of its column. This is the default behaviour.

• Impute the missing attribute by building a separate model which will attempt to predict the missing attribute based on the the attributes that are available.

The second option is what we refer to as the Dynamic Imputer. The online nature of SymetryML enables these imputation models to be build on the fly as the prediction file is being processed. Enabling this functionality is as simple as specifying impute=true in the query parameter of the Predict APIs.

Sample Request/Response DS Predict

Request:
POST url="http://charm:8080/symetry/rest/c1/projects/irisTest/dss/predict/testLDA?indsname=dsin&outdsname=dsout&impute=true
Request Body:
{
  "attributeNames":["sepal_length","sepal_width","petal_length","petal_width"],
  "attributeTypes":["C","C","C","C"]
}

Response Header:
Location: http://charm:8080/symetry/rest/c1/jobs/2

Response:
{"statusCode":"ACCEPTED","statusString":"Job Created","values":{}}

Job Request:
GET url="http://charm:8080/symetry/rest/c1/jobs/2

Job Response:
{"statusCode":"OK","statusString":"Job is finished","values":{"dataframe":{"attributeNames":["resZ_Result","res_Result","normZ_Result","sepal_length","sepal_width","petal_length","petal_width","sepal_lengt_b1","sepal_lengt_b2","sepal_width_b1","sepal_width_b2","petal_length_b1","petal_length_b2","petal_width_b1","petal_width_b2","Iris_setosa","Iris_versicolor","Iris_virginica"],"data":[["-0.4559922893809558","1","0.2929126883636288","4.3","3","1.1","0.1","1","0","0","1","1","0","1","0","1","0","0"],["-0.5225734165408404","1","0.2965158290613048","

(…)

["-2.6526076600477295","1","0.4232667751179995","6.9","3.1","5.4","2.1","0","1","0","1","0","1","0","1","0","0","1"]]}}}

Sample Request/Response Predict

Request:
POST url="http://charm:8080/symetry/rest/c1/projects/irisTest/predict/testLDA?impute=true"

Body:
{"attributeNames":["petal_length","sepal_width"],"data":[["1.5","5"]],"attributeTypes":["C","C"]}

Response:
{"statusCode":"OK","statusString":"OK","values":{"KSVSMap":{"values":[{"res":"1","normZ":"0.40997859439073814","resZ":"-2.4426003268943006","z0":"-3.840891864408894","z1LTz0":"false"}]}}}

Prediction REST API Canonical URL

POST /symetry/rest/{cid}/projects/{pid}/predict/{modelid} [body=DataFrame]

HTTP Responses

HTTP Status Code	HTTP Status Message	Description
200	OK	Success.
400	BAD REQUEST	Invalid RTM ID. {"statusCode":"BAD_REQUEST","statusString":"Cannot Find SYMETRYML id[r2] for Customer id [c1]","values":{}}
500	INTERNAL SERVER ERROR	Invalid model ID. {"statusCode":"INTERNAL_SERVER_ERROR","statusString":"Generic Exception \u003cError no such model \u003cm3\u003e\u003e","values":{}}

HTTP Response Entity

HTTP Response Entity	Example
KSVSMap	Contain information about the prediction.

Sample Request Response

Request:
POST url="http://charm:8080/symetry/rest/c1/projects/irisTest/predict/testLDA"

Body:
{"attributeNames":["petal_length","sepal_width"],"data":[["1.5","5"]],"attributeTypes":["C","C"]}

Response:
{"statusCode":"OK","statusString":"OK","values":{"KSVSMap":{"values":[{"res":"1","normZ":"0.40997859439073814","resZ":"-2.4426003268943006","z0":"-3.840891864408894","z1LTz0":"false"}]}}}

Model Prediction Map Keys

The prediction service returns a KSVSMap. This map is an array of map<key, value>, with key and value being strings. This is different from KSVDMap, where the values are double. The following table lists the model types and their corresponding KSVDMap key. The following table enumerates the keys returned by each model.

Model Types	Returned Keys
ECOD	res:1 or 0 (1 if ecodscore > chosen threshold score) ecodscore: the anomaly score of the ECOD model
EVT	evt_is_anomaly: 1 or 0 (1 if score > threshold) evt_threshold: value of threshold if EVT is one sided evt_threshold_upper: value of the upper threshold if EVT is double sided evt_threshold_lower: value of the lower threshold if EVT is double sided
Partial Least Square Regression	If OneQ is true q Else 1 … q predictions
Principal Component Regression	If pcr_type == full return res. Otherwise, If OneQ is true return q Else 1 … q
LDA & MLDA	- res, 1 or 0 - resZ - z0 - z1GTz0 Examples: res :1 resZ :-1.66850 z0 :0.240482 z1GTz0 :true LDA also return resU, which is similar to res but use a different threshold for when the positive and negative examples are unbalanced in the training data.
LSVM	- res, 1 or 0
LSVR	- res, the regression result
MLR	- res, the regression result
QDA	For each target in the model, a key,value pair is returned. The format of the key is ‘<attribute name>’. The value is the number computed by the model for that target. Example: Corn :12.0509 * *Sugarbeets :11.4417 * *Clover :11.7757 * *Soybeans :12.0040 Cotton :12.0634 The best scoring target is also returned in the ‘res’ key.
Bayes	res = 0 or 1. Prob = contains the likelihood of the result.
Markov Chains	next = contains the next state next.p = contains probability of the next state seq.like = contains the likelihood of the input sequence
Hidden Markov Model	seq = sequence of hidden states
Random Forest Classifier	-res, the class id with the highest probability. The probability for each classes is also returned. For a model with 3 classes their ids would be [0, 1, 2].
Kaplan Meier	Prediction is not defined for Kaplan Meier model. Predict should not be called on models of type KM.