odps/static/algorithms/regression.xml

<?xml version='1.0' encoding='UTF-8'?>
<algorithms baseClass="BaseTrainingAlgorithm">
    <algorithm codeName="LinearRegression">
        <reloadFields>false</reloadFields>
        <docs><![CDATA[
        将输入表的数据，选择某些列作为feature，某一列作为label，进行线性回归训练，产生线性回归模型

        %params%
        ]]></docs>
        <params>
            <param name="inputTableName">
                <exporter>get_input_table_name</exporter>
                <inputName>input</inputName>
            </param>
            <param name="featureColNames">
                <alias>featureCols</alias>
                <exporter>get_feature_columns</exporter>
                <inputName>input</inputName>
            </param>
            <param name="labelColName">
                <alias>labelCol</alias>
                <exporter>get_label_column</exporter>
                <inputName>input</inputName>
            </param>
            <param name="modelName">
                <exporter>generate_model_name</exporter>
                <outputName>output</outputName>
            </param>
            <param name="inputTablePartitions">
                <exporter>get_input_partitions</exporter>
                <inputName>input</inputName>
            </param>
            <param name="maxIter">
                <value>100</value>
                <docs>最大迭代次数</docs>
            </param>
            <param name="epsilon">
                <value>0.000001</value>
                <docs>终止条件，即两次迭代之间log-likelihood的差</docs>
            </param>
            <param name="regularizedType">
                <docs>正则化类型，可选 'l1'、'l2'和 None</docs>
            </param>
            <param name="regularizedLevel">
                <value>1.0</value>
                <min>0</min>
                <docs>正则项系数,当 regularized_type 为 None 时，该参数会被忽略</docs>
            </param>
            <param name="itemDelimiter">
                <exporter>get_item_delimiter</exporter>
                <inputName>input</inputName>
            </param>
            <param name="kvDelimiter">
                <exporter>get_kv_delimiter</exporter>
                <inputName>input</inputName>
            </param>
            <param name="enableSparse">
                <exporter>get_enable_sparse</exporter>
                <inputName>input</inputName>
            </param>
        </params>
        <ports>
            <port name="input">
                <ioType>INPUT</ioType>
                <sequence>1</sequence>
                <type>DATA</type>
                <docs>train set with defined features and labels</docs>
            </port>
            <port name="output">
                <ioType>OUTPUT</ioType>
                <sequence>1</sequence>
                <type>MODEL</type>
                <docs>trained model which can be used in prediction</docs>
            </port>
        </ports>
        <metas>
            <meta name="xflowName" value="linearregression"/>
            <meta name="xflowProjectName" value="algo_public"/>
        </metas>
    </algorithm>
    <algorithm codeName="xgboost">
        <docs><![CDATA[
        xgboost是一种高效率boosting算法，适用回归和二分类问题，详见 https://github.com/dmlc/xgboost 。
        PAI平台目前只开放树形模型（gbtree），暂不开放线性模型（gblinear）。

        %params%
        ]]></docs>
        <reloadFields>false</reloadFields>
        <params>
            <param name="featureColNames">
                <alias>featureCols</alias>
                <exporter>get_feature_columns</exporter>
                <inputName>input</inputName>
                <docs>输入表中用于训练的特征的列名</docs>
            </param>
            <param name="labelColName">
                <alias>labelCol</alias>
                <exporter>get_label_column</exporter>
                <inputName>input</inputName>
                <docs>输入表中标签列的列名</docs>
            </param>
            <param name="booster">
                <value>gbtree</value>
            </param>
            <param name="inputTableName">
                <exporter>get_input_table_name</exporter>
                <inputName>input</inputName>
            </param>
            <param name="inputPartitions">
                <exporter>get_input_partitions</exporter>
                <inputName>input</inputName>
            </param>
            <param name="modelName">
                <exporter>generate_model_name</exporter>
                <outputName>output</outputName>
                <docs>输出的模型名</docs>
            </param>
            <param name="itemDelimiter">
                <exporter>get_item_delimiter</exporter>
                <inputName>input</inputName>
                <docs>item(key value对)之间的分隔符</docs>
            </param>
            <param name="kvDelimiter">
                <exporter>get_kv_delimiter</exporter>
                <inputName>input</inputName>
                <docs>表中每个item的key和value之间的分隔符</docs>
            </param>
            <param name="enableSparse">
                <exporter>get_enable_sparse</exporter>
                <inputName>input</inputName>
                <docs>是否稀疏数据</docs>
            </param>
            <param name="silent">
                <value>1</value>
            </param>
            <param name="eta">
                <value>0.3</value>
                <min>0</min>
                <max>1</max>
                <docs>为了防止过拟合，更新过程中用到的收缩步长。在每次提升计算之后，算法会直接获得新特征的权重。 eta通过缩减特征的权重使提升计算过程更加保守</docs>
            </param>
            <param name="gamma">
                <value>0</value>
                <min>0</min>
                <docs>最小损失衰减</docs>
            </param>
            <param name="max_depth">
                <value>6</value>
                <min>0</min>
                <docs>树的最大深度</docs>
            </param>
            <param name="min_child_weight">
                <value>1</value>
                <min>0</min>
                <docs>孩子节点中最小的样本权重和。如果一个叶子节点的样本权重和小于min_child_weight则拆分过程结束。在线性回归模型中，这个参数是指建立每个模型所需的最小样本数</docs>
            </param>
            <param name="max_delta_step">
                <value>0</value>
                <min>0</min>
                <docs>每个树所允许的最大delta步进</docs>
            </param>
            <param name="colsample_bytree">
                <value>1</value>
                <min>0</min>
                <max>1</max>
                <docs>在建立树时对特征采样的比例</docs>
            </param>
            <param name="base_score">
                <value>0.5</value>
                <docs>初始的predict score</docs>
            </param>
            <param name="seed">
                <value>0</value>
                <docs>随机数的种子</docs>
            </param>
            <param name="objective">
                <value>reg:linear</value>
                <docs>定义学习任务及相应的学习目标，可选的目标函数</docs>
            </param>
            <param name="eval_metric">
                <value>error</value>
            </param>
        </params>
        <ports>
            <port name="input">
                <ioType>INPUT</ioType>
                <sequence>1</sequence>
                <type>DATA</type>
                <docs>train set with defined features and labels</docs>
            </port>
            <port name="output">
                <ioType>OUTPUT</ioType>
                <sequence>1</sequence>
                <type>MODEL</type>
                <docs>trained model which can be used in prediction</docs>
            </port>
        </ports>
        <metas>
            <meta name="xflowName" value="xgboost"/>
            <meta name="xflowProjectName" value="algo_public"/>
        </metas>
    </algorithm>
    <algorithm codeName="GBDT">
        <reloadFields>false</reloadFields>
        <docs><![CDATA[
        Gradient Boosting Decision Tree is a machine learning technique for regression problems, which produces a prediction
        model in the form of an ensemble of weak prediction models, typically decision trees. It builds the model in
        a stage-wise fashion like other boosting methods do, and it generalizes them by allowing optimization of an arbitrary
        differentiable loss function. The gradient boosting method can also be used for classification problems by reducing them
        to regression with a suitable loss function.

        GBDT could be used in all regression problems (linear/ nonlinear) while logistic regression can only be used for linear
        regression. So the application of GBDT is very wide.

        %params%

        :Example:
        >>> labeled = ds.label_field('class')
        >>> model = GBDT(tree_count=500).train(labeled)
        >>> predicted = model.predict(dataset_to_be_predicted)
        >>> predicted.store_odps('predicted')
        ]]></docs>
        <params>
            <param name="inputTableName" required="true">
                <exporter>get_input_table_name</exporter>
                <inputName>input</inputName>
            </param>
            <param name="inputTablePartitions">
                <exporter>get_input_partitions</exporter>
                <inputName>input</inputName>
            </param>
            <param name="modelName" required="true">
                <exporter>generate_model_name</exporter>
                <outputName>output</outputName>
                <docs>name of modle to be output</docs>
            </param>
            <param name="labelColName" required="true">
                <alias>labelCol</alias>
                <exporter>get_label_column</exporter>
                <inputName>input</inputName>
                <docs>name of label column selected from input table</docs>
            </param>
            <param name="featureColNames" required="true">
                <alias>featureCols</alias>
                <exporter>get_feature_columns</exporter>
                <inputName>input</inputName>
                <docs>feature columns of input table, used for training</docs>
            </param>
            <param name="groupIDColName">
                <exporter>get_group_id_column</exporter>
                <inputName>input</inputName>
                <docs>name of grouping column. The default is to take a whole table as a group</docs>
            </param>
            <param name="metricType">
                <value>2</value>
            </param>
            <param name="treeCount">
                <value>500</value>
                <min>1</min>
                <max>10000</max>
                <docs>num. of trees.the range is [1,10000] and the default value is 500</docs>
            </param>
            <param name="shrinkage">
                <value>0.05</value>
                <min>0</min>
                <max>1</max>
                <docs>learning rate: the range is (0, 1] and default value is 0.05</docs>
            </param>
            <param name="maxLeafCount">
                <value>32</value>
                <min>2</min>
                <max>1000</max>
                <docs>maximal num. of leaves: it must be an integer. The range is [2,1000] and the default value is 32
                </docs>
            </param>
            <param name="maxDepth">
                <value>11</value>
                <min>1</min>
                <max>11</max>
                <docs>maximal depth of tree.it must be an integer. The range is [1, 11] and the default value is 11
                </docs>
            </param>
            <param name="minLeafSampleCount">
                <value>500</value>
                <docs>minimal leaf size: it must be an integer. The range is [100,1000] and the default value is 500
                </docs>
            </param>
            <param name="sampleRatio">
                <value>0.6</value>
                <min>0</min>
                <max>1</max>
                <docs>ratio of instances for training. the range is (0,1] and the default value is 0.6</docs>
            </param>
            <param name="featureRatio">
                <value>0.6</value>
                <min>0</min>
                <max>1</max>
                <docs>ratio of features for training: the range is (0,1] and the default value is 0.6</docs>
            </param>
            <param name="testRatio">
                <value>0.0</value>
                <min>0</min>
                <max>1</max>
                <docs>ratio of test instances. the range is [0,1) and the default value is 0.0</docs>
            </param>
            <param name="tau">
                <value>0.6</value>
                <min>0</min>
                <max>1</max>
            </param>
            <param name="p">
                <value>1</value>
                <min>1</min>
                <max>5</max>
            </param>
            <param name="randSeed">
                <value>0</value>
                <min>0</min>
                <max>10</max>
                <docs>random seed. It must be an integer and the range is [0,10]. The default value is 0</docs>
            </param>
            <param name="newtonStep">
                <value>0</value>
                <min>0</min>
                <max>1</max>
            </param>
            <param name="featureSplitValueMaxSize">
                <value>500</value>
                <min>1</min>
                <max>1000</max>
                <docs>maximal splits: the range is [1,1000] and the default value is 500</docs>
            </param>
            <param name="lossType">
                <value>3</value>
                <docs><![CDATA[function: the default type is regression loss. If the grouping column is not set, you can
only select regression loss. Select the types:
GBRANK LOSS `thesis <http://www.cc.gatech.edu/~zha/papers/fp086-zheng.pdf>`_,
DCG LOSS `thesis <http://research-srv.microsoft.com/pubs/132652/MSR-TR-2010-82.pdf>`_,
NDCG LOSS `thesis <http://research-srv.microsoft.com/pubs/132652/MSR-TR-2010-82.pdf>`_,
REGRESSION LOSS mean squared error loss]]></docs>
            </param>
        </params>
        <ports>
            <port name="input">
                <ioType>INPUT</ioType>
                <sequence>1</sequence>
                <type>DATA</type>
                <docs>train set with defined features and labels</docs>
            </port>
            <port name="output">
                <ioType>OUTPUT</ioType>
                <sequence>1</sequence>
                <type>MODEL</type>
                <docs>trained model which can be used in prediction</docs>
            </port>
        </ports>
        <metas>
            <meta name="xflowName" value="GBDT"/>
            <meta name="xflowProjectName" value="algo_public"/>
        </metas>
    </algorithm>
</algorithms>