Need your help building Neural Network using a nested table

Question

Greetings,

Thanks for the instructions for I am too a DM beginner. Having accumulated some expertise in SSAS development area, I am now tackling a similar issue to http://social.msdn.microsoft.com/Forums/en-US/sqldatamining/thread/7da50abd-cab2-4ee0-bea7-a14769c9312b but need to build a neural network, train it and predict. I also need to use nested tables:

CASE table:

CaseID (PK)

CASE Name (input?)

ShouldAlert (predict)

NESTED table

Index (PK)

CaseID (FK)

Value

We are trying to build a prediction system trained on sets of arrays to generate alerts on graphs. I have built the NN model and set Value as the key. Trying to train the model to predict on ShouldAlert column based on Value. When I am in Mining Model Viewer, it fails with the message:

Execution of the managed stored procedure GetAttributeScores failed with the following error: Exception has been thrown by the target of an invocation.The value '5.9999999999999995E-4' is not found for the 'Value' node..

Any assistance would be appreciated. Thanks!

Answer 1

This looks like a bug to me. Can you create a back up file with your model and share it?

---

Tatyana Yakushev [PredixionSoftware.com]

Answer 2

Thanks for your assistance Tatyana. The backup file is stored in http://www.4shared.com/folder/uh1PjH9o/_online.html?refurl=sharedfolderurl

I am looking for general directions as well. My first approach was to create a single table with joining case table and nested table on caseid key. The new table had 2 keys: caseid and index. I was trying to predict ShouldAlert based on Value and was able to achieve pretty good results. Then I tried to evaluate the model based on test table with values and failed getting all TRUES. I believe I incorrectly joined the model with the evaluation sets.

Then I learned about the nested tables and decided to try. Would you also provide the joins I need to use to evaluate test data? Fields in the evaluation table are named the same as in the nested table. Thanks a lot!

Answer 3

I can reproduce the issue but I don't know why store procedure throws exception because I don't have a source code. May be somebody from Microsoft can debug the issue.

I don't understand the data and what you are trying to model to recommend a good approach for creating a model. It is possible that the first model you created didn't find any patterns and was always predicting the more popular state. You can start with creating a decision tree model, it is easier to understand.

---

Tatyana Yakushev [PredixionSoftware.com]

Answer 4

Thanks Tatyana,

The concept behind the model is easy. The case table contains the ids of all the 'graphs' along with their associated 'ShouldAlert' prediction field. The Mvr Alert Training Data Point nested table contains point values for each of the linear graphs. The task is to train the model to predict whether certain pattern should alert or not. Usually the alert should go off once the points experience a deviation: 10x drop/increase of the mean etc.

As a novice, I am looking for any guidance to change/re-arrange, rebuild the model etc. If anyone has ideas please share. Thanks

Answer 5

Are values in "Mvr Alert Training Data Point" measurements of the same thing at different points of time?

---

Tatyana Yakushev [PredixionSoftware.com]

Answer 6

Yes, you can think of Index field as different days

Answer 7

The reason I'm asking is I am trying to understand what algorithm might give you good predictions. For example, may be you should use time series algorithm and alert when actual value is very different from prediction.

---

Tatyana Yakushev [PredixionSoftware.com]

Answer 8

Thanks. Time Series is not exactly what I was looking for. Here is another anology.  We have a bunch of cases, each representing an array of values with some anomalies to predict - 500 point population, 10x spike at the end or 1000 point population with 100x drop in the beginning etc. We load all these cases into the data mining machine to train it to predict similar cases. Basically, when the DM finds a similar real-life case - 10x spike at the end, for instance, the alert should light up.  We want to build a graph alert system this way and would like to look and evaluate the case (array) as a whole deciding whether to alert or not.