Options
Trouble with Optimizing a Model
Contributor I
Dear RM Community,
I am currently engaged in a project to better understand account retention. I would like to create a decision tree model that intelligently categorizes the types of customers that close their accounts after 12 months. This will hopefully give us a better look at which customers we might need to reach out to in order to retain their business. However, along the way, I have hit some bumps in the road with my coding and am in need of some help. I have heard great things about this community and its helpfulness, so hopefully we can work together to push past this road block.
My issues, broken down into three segments, lead into one another and are as follows:
Optimized Parameters
• Using the Optimize Parameters (Evolutionary) process, I was able to determine the optimal parameters to maximize weighted mean recall and weighted mean precision
• Issue: I want weighted mean recall and weighted mean precision to be optimized at the same time and not have a main criterion (not a big issue if this cannot be done so long as both weighted means are significant)
Replication
• After I have determine the optimized parameters, I then attempted to recreate the same model that produced the optimized results by using the optimized parameters in a new decision tree
o To my understanding, theoretically this should produce a similar tree as well as similar results but will vary because the training set will be slightly different
o Results and tree change drastically for the worse
• Issue: I want to be able to find an optimized model and be able to replicate/create a new model for future use
Accuracy
• Now that I have successfully partially replicated the optimized model and produced the performance of the model, the overall strength of the model has decreased
• Issue: the model needs to be able to be replicated and perform just as well as when it was optimized
Thank you for any and all help. I greatly appreciate it.
Cheers,
Jerrod
I am currently engaged in a project to better understand account retention. I would like to create a decision tree model that intelligently categorizes the types of customers that close their accounts after 12 months. This will hopefully give us a better look at which customers we might need to reach out to in order to retain their business. However, along the way, I have hit some bumps in the road with my coding and am in need of some help. I have heard great things about this community and its helpfulness, so hopefully we can work together to push past this road block.
My issues, broken down into three segments, lead into one another and are as follows:
Optimized Parameters
• Using the Optimize Parameters (Evolutionary) process, I was able to determine the optimal parameters to maximize weighted mean recall and weighted mean precision
• Issue: I want weighted mean recall and weighted mean precision to be optimized at the same time and not have a main criterion (not a big issue if this cannot be done so long as both weighted means are significant)
Replication
• After I have determine the optimized parameters, I then attempted to recreate the same model that produced the optimized results by using the optimized parameters in a new decision tree
o To my understanding, theoretically this should produce a similar tree as well as similar results but will vary because the training set will be slightly different
o Results and tree change drastically for the worse
• Issue: I want to be able to find an optimized model and be able to replicate/create a new model for future use
Accuracy
• Now that I have successfully partially replicated the optimized model and produced the performance of the model, the overall strength of the model has decreased
• Issue: the model needs to be able to be replicated and perform just as well as when it was optimized
Thank you for any and all help. I greatly appreciate it.
Cheers,
Jerrod
<?xml version="1.0" encoding="UTF-8" standalone="no"?>
<process version="6.3.000">
<context>
<input/>
<output/>
<macros/>
</context>
<operator activated="true" class="process" compatibility="6.3.000" expanded="true" name="Process">
<process expanded="true">
<operator activated="true" class="read_csv" compatibility="6.3.000" expanded="true" height="60" name="Read CSV" width="90" x="112" y="30">
<parameter key="csv_file" value="Z:\Marketing\Staff\Jerrod Johnson\MOB18_v2.csv"/>
<parameter key="column_separators" value=","/>
<parameter key="first_row_as_names" value="false"/>
<list key="annotations">
<parameter key="0" value="Name"/>
</list>
<parameter key="encoding" value="windows-1252"/>
<list key="data_set_meta_data_information">
<parameter key="0" value="Region.true.polynominal.attribute"/>
<parameter key="1" value="Line Of Business.true.polynominal.attribute"/>
<parameter key="2" value="Company Type.true.polynominal.attribute"/>
<parameter key="3" value="Revenue (US Dollars; million).true.real.attribute"/>
<parameter key="4" value="Total Employees.true.real.attribute"/>
<parameter key="5" value="NAICS DESC (JH).true.polynominal.attribute"/>
<parameter key="6" value="ProductClass.true.polynominal.attribute"/>
<parameter key="7" value="StatusOpen.true.real.attribute"/>
</list>
</operator>
<operator activated="true" class="select_attributes" compatibility="6.3.000" expanded="true" height="76" name="Select Attributes" width="90" x="112" y="120">
<parameter key="attribute_filter_type" value="subset"/>
<parameter key="attributes" value="Company Type|NAICS DESC (JH)|Region|Revenue (US Dollars; million)|StatusOpen|Total Employees|Line Of Business|ProductClass"/>
</operator>
<operator activated="true" class="set_role" compatibility="6.3.000" expanded="true" height="76" name="Set Role" width="90" x="112" y="255">
<parameter key="attribute_name" value="StatusOpen"/>
<parameter key="target_role" value="label"/>
<list key="set_additional_roles"/>
</operator>
<operator activated="true" class="numerical_to_polynominal" compatibility="6.3.000" expanded="true" height="76" name="Numerical to Polynominal" width="90" x="112" y="390">
<parameter key="attribute_filter_type" value="subset"/>
<parameter key="attributes" value="StatusOpen"/>
<parameter key="include_special_attributes" value="true"/>
</operator>
<operator activated="true" class="optimize_parameters_evolutionary" compatibility="6.3.000" expanded="true" height="112" name="Optimize Parameters (Evolutionary)" width="90" x="380" y="165">
<list key="parameters">
<parameter key="Learner Evo.maximal_depth" value="[1;20]"/>
<parameter key="Learner Evo.confidence" value="[1.0E-7;0.5]"/>
<parameter key="Learner Evo.minimal_gain" value="[0.005;.15]"/>
<parameter key="Learner Evo.minimal_leaf_size" value="[1;8]"/>
<parameter key="Learner Evo.minimal_size_for_split" value="[1.0;6]"/>
<parameter key="Learner Evo.number_of_prepruning_alternatives" value="[1;6]"/>
<parameter key="Validation Evo.split_ratio" value="[0.6;.9]"/>
</list>
<parameter key="max_generations" value="100"/>
<parameter key="population_size" value="50"/>
<parameter key="use_local_random_seed" value="true"/>
<parameter key="show_convergence_plot" value="true"/>
<process expanded="true">
<operator activated="true" class="split_validation" compatibility="6.3.000" expanded="true" height="112" name="Validation Evo" width="90" x="447" y="30">
<parameter key="split_ratio" value="0.8931198308911473"/>
<process expanded="true">
<operator activated="true" class="parallel_decision_tree" compatibility="6.3.000" expanded="true" height="76" name="Learner Evo" width="90" x="179" y="30">
<parameter key="criterion" value="information_gain"/>
<parameter key="maximal_depth" value="14"/>
<parameter key="confidence" value="0.06462070743893004"/>
<parameter key="minimal_gain" value="0.005"/>
<parameter key="minimal_leaf_size" value="4"/>
<parameter key="minimal_size_for_split" value="3"/>
<parameter key="number_of_prepruning_alternatives" value="6"/>
</operator>
<connect from_port="training" to_op="Learner Evo" to_port="training set"/>
<connect from_op="Learner Evo" from_port="model" to_port="model"/>
<portSpacing port="source_training" spacing="0"/>
<portSpacing port="sink_model" spacing="0"/>
<portSpacing port="sink_through 1" spacing="0"/>
</process>
<process expanded="true">
<operator activated="true" class="apply_model" compatibility="6.3.000" expanded="true" height="76" name="Apply Model (6)" width="90" x="112" y="30">
<list key="application_parameters"/>
</operator>
<operator activated="true" class="performance_classification" compatibility="6.3.000" expanded="true" height="76" name="Performance (6)" width="90" x="246" y="30">
<parameter key="accuracy" value="false"/>
<parameter key="weighted_mean_recall" value="true"/>
<parameter key="weighted_mean_precision" value="true"/>
<list key="class_weights"/>
</operator>
<connect from_port="model" to_op="Apply Model (6)" to_port="model"/>
<connect from_port="test set" to_op="Apply Model (6)" to_port="unlabelled data"/>
<connect from_op="Apply Model (6)" from_port="labelled data" to_op="Performance (6)" to_port="labelled data"/>
<connect from_op="Performance (6)" from_port="performance" to_port="averagable 1"/>
<portSpacing port="source_model" spacing="0"/>
<portSpacing port="source_test set" spacing="0"/>
<portSpacing port="source_through 1" spacing="0"/>
<portSpacing port="sink_averagable 1" spacing="0"/>
<portSpacing port="sink_averagable 2" spacing="0"/>
</process>
</operator>
<connect from_port="input 1" to_op="Validation Evo" to_port="training"/>
<connect from_op="Validation Evo" from_port="model" to_port="result 1"/>
<connect from_op="Validation Evo" from_port="averagable 1" to_port="performance"/>
<portSpacing port="source_input 1" spacing="0"/>
<portSpacing port="source_input 2" spacing="0"/>
<portSpacing port="sink_performance" spacing="0"/>
<portSpacing port="sink_result 1" spacing="0"/>
<portSpacing port="sink_result 2" spacing="0"/>
</process>
</operator>
<operator activated="true" class="set_parameters" compatibility="6.3.000" expanded="true" height="76" name="Set Parameters" width="90" x="514" y="75">
<list key="name_map">
<parameter key="Learner Evo" value="OptimalLearner"/>
<parameter key="Validation Evo" value="OptimValidation"/>
</list>
</operator>
<operator activated="true" class="read_csv" compatibility="6.3.000" expanded="true" height="60" name="Read CSV (3)" width="90" x="380" y="570">
<parameter key="csv_file" value="Z:\Marketing\Staff\Jerrod Johnson\MOB18_v2.csv"/>
<parameter key="column_separators" value=","/>
<parameter key="first_row_as_names" value="false"/>
<list key="annotations">
<parameter key="0" value="Name"/>
</list>
<parameter key="encoding" value="windows-1252"/>
<list key="data_set_meta_data_information">
<parameter key="0" value="Region.true.polynominal.attribute"/>
<parameter key="1" value="Line Of Business.true.polynominal.attribute"/>
<parameter key="2" value="Company Type.true.polynominal.attribute"/>
<parameter key="3" value="Revenue (US Dollars; million).true.real.attribute"/>
<parameter key="4" value="Total Employees.true.real.attribute"/>
<parameter key="5" value="NAICS DESC (JH).true.polynominal.attribute"/>
<parameter key="6" value="ProductClass.true.polynominal.attribute"/>
<parameter key="7" value="StatusOpen.true.real.attribute"/>
</list>
</operator>
<operator activated="true" class="select_attributes" compatibility="6.3.000" expanded="true" height="76" name="Select Attributes (3)" width="90" x="380" y="480">
<parameter key="attribute_filter_type" value="subset"/>
<parameter key="attributes" value="Company Type|NAICS DESC (JH)|Region|Revenue (US Dollars; million)|StatusOpen|Total Employees|Line Of Business|ProductClass"/>
</operator>
<operator activated="true" class="set_role" compatibility="6.3.000" expanded="true" height="76" name="Set Role (2)" width="90" x="380" y="390">
<parameter key="attribute_name" value="StatusOpen"/>
<parameter key="target_role" value="label"/>
<list key="set_additional_roles"/>
</operator>
<operator activated="true" class="numerical_to_polynominal" compatibility="6.3.000" expanded="true" height="76" name="Numerical to Polynominal (2)" width="90" x="380" y="300">
<parameter key="attribute_filter_type" value="subset"/>
<parameter key="attributes" value="StatusOpen"/>
<parameter key="include_special_attributes" value="true"/>
</operator>
<operator activated="true" class="split_validation" compatibility="6.3.000" expanded="true" height="112" name="OptimValidation" width="90" x="514" y="300">
<parameter key="split_ratio" value="0.8931198308911473"/>
<process expanded="true">
<operator activated="true" class="parallel_decision_tree" compatibility="6.3.000" expanded="true" height="76" name="OptimalLearner" width="90" x="179" y="30">
<parameter key="maximal_depth" value="14"/>
<parameter key="confidence" value="0.06462070743893004"/>
<parameter key="minimal_gain" value="0.005"/>
<parameter key="minimal_leaf_size" value="4"/>
<parameter key="minimal_size_for_split" value="3"/>
<parameter key="number_of_prepruning_alternatives" value="6"/>
</operator>
<connect from_port="training" to_op="OptimalLearner" to_port="training set"/>
<connect from_op="OptimalLearner" from_port="model" to_port="model"/>
<portSpacing port="source_training" spacing="0"/>
<portSpacing port="sink_model" spacing="0"/>
<portSpacing port="sink_through 1" spacing="0"/>
</process>
<process expanded="true">
<operator activated="true" class="apply_model" compatibility="6.3.000" expanded="true" height="76" name="Apply Model (12)" width="90" x="45" y="30">
<list key="application_parameters"/>
</operator>
<operator activated="true" class="performance_classification" compatibility="6.3.000" expanded="true" height="76" name="Performance (12)" width="90" x="313" y="30">
<parameter key="accuracy" value="false"/>
<parameter key="weighted_mean_recall" value="true"/>
<parameter key="weighted_mean_precision" value="true"/>
<list key="class_weights"/>
</operator>
<connect from_port="model" to_op="Apply Model (12)" to_port="model"/>
<connect from_port="test set" to_op="Apply Model (12)" to_port="unlabelled data"/>
<connect from_op="Apply Model (12)" from_port="labelled data" to_op="Performance (12)" to_port="labelled data"/>
<connect from_op="Performance (12)" from_port="performance" to_port="averagable 1"/>
<portSpacing port="source_model" spacing="0"/>
<portSpacing port="source_test set" spacing="0"/>
<portSpacing port="source_through 1" spacing="0"/>
<portSpacing port="sink_averagable 1" spacing="0"/>
<portSpacing port="sink_averagable 2" spacing="0"/>
</process>
</operator>
<connect from_op="Read CSV" from_port="output" to_op="Select Attributes" to_port="example set input"/>
<connect from_op="Select Attributes" from_port="example set output" to_op="Set Role" to_port="example set input"/>
<connect from_op="Set Role" from_port="example set output" to_op="Numerical to Polynominal" to_port="example set input"/>
<connect from_op="Numerical to Polynominal" from_port="example set output" to_op="Optimize Parameters (Evolutionary)" to_port="input 1"/>
<connect from_op="Optimize Parameters (Evolutionary)" from_port="performance" to_port="result 4"/>
<connect from_op="Optimize Parameters (Evolutionary)" from_port="parameter" to_op="Set Parameters" to_port="parameter set"/>
<connect from_op="Optimize Parameters (Evolutionary)" from_port="result 1" to_port="result 6"/>
<connect from_op="Set Parameters" from_port="parameter set" to_port="result 5"/>
<connect from_op="Read CSV (3)" from_port="output" to_op="Select Attributes (3)" to_port="example set input"/>
<connect from_op="Select Attributes (3)" from_port="example set output" to_op="Set Role (2)" to_port="example set input"/>
<connect from_op="Set Role (2)" from_port="example set output" to_op="Numerical to Polynominal (2)" to_port="example set input"/>
<connect from_op="Numerical to Polynominal (2)" from_port="example set output" to_op="OptimValidation" to_port="training"/>
<connect from_op="OptimValidation" from_port="model" to_port="result 1"/>
<connect from_op="OptimValidation" from_port="training" to_port="result 2"/>
<connect from_op="OptimValidation" from_port="averagable 1" to_port="result 3"/>
<portSpacing port="source_input 1" spacing="0"/>
<portSpacing port="sink_result 1" spacing="0"/>
<portSpacing port="sink_result 2" spacing="0"/>
<portSpacing port="sink_result 3" spacing="0"/>
<portSpacing port="sink_result 4" spacing="0"/>
<portSpacing port="sink_result 5" spacing="0"/>
<portSpacing port="sink_result 6" spacing="0"/>
<portSpacing port="sink_result 7" spacing="0"/>
</process>
</operator>
</process>
0
Answers
The most straight forward way is to the Combine Performances Operator. Then you can optimize on the weighted performance. Another way would be to use the Log operator and Log the performances in each step. Than you can choose by hand what’s your prefered setting.
Furthermore you can of course use example weights in the calculation of performance values. So you could use accuracy and weight by hand.
After I have determine the optimized parameters, I then attempted to recreate the same model that produced the optimized results by using the optimized parameters in a new decision tree *
The issue might be related to the decision tree itself. The decision tree might be volatile to a small change. If the root node just changes a bit, thus the resulting decision tree might change drastically.
Could you comment on the used validation? How many examples did you use? Is the Decision Tree set as a Model? A Random Forest should be less volatile.
This issue partially stems from my lack of understanding of how the Optimize Parameters process actually works *
The evolutionary algorithm is kind of complex. Have you tried to use Optimize (Grid) with a bit of a wide grid? If you have less than 5 parameters it should be work fine.
Now that I have successfully partially replicated the optimized model and produced the performance of the model, the overall strength of the model has decreased *
It might be that you overtrained your algorithm by choosing the correct parameters which suits this special subset of data. Again i would like to have a look on that process and it’s validation.
Edit: I have looked on your process. try to add a validation around your optimiziation and look if the resulting parameters are similar.
Cheers,
Martin
Dortmund, Germany
I will try to work through this process. Seems very interesting.
Could you comment on the used validation? How many examples did you use?
If I understand your question correctly, I have been using split validation for everything. After optimizing, it returns an optimal split of .79. I have 603 examples so it uses around 482 examples for the training set.
A Random Forest should be less volatile.
I am not set on using a decision tree, but I tried the Random Forest model and it seemed to perform the same.
Have you tried to use Optimize (Grid) with a bit of a wide grid?
Because I am unsure about any of the parameters within the Decision Tree, I would have to optimize on every parameter of the decision tree in the Optimize (Grid). I have tried this and it took quite a while but returned a model that when replicated, performed worse than my current model. I will try using the grid again.
try to add a validation around your optimization
Whenever I add a validation around the optimization, it says it is lacking a model output. Does the Optimize Parameters output a model? I don’t believe that it has ever output a model for me.
Thank you for all of your help!
Best,
Jerrod
Cheers,
Martin
Dortmund, Germany