Similarity solutions
Summarize
Summary of Similarity solutions
Similarity solutions leverage Machine Learning (ML) to compare text between resolved and open alert records, enabling the reuse of resolution approaches for similar alerts. This capability helps streamline incident management by identifying relevant past resolutions based on textual similarity.
Show less
Training a similarity solution
Training involves compiling a collection of words from key fields—such as Short Description, Description, Source, Type, Resource, and Metric Name—in resolved alerts. ML uses this collection to match against open alerts to find similarities. Effective training requires:
- A filter that returns at least one relevant record; ideally between 30,000 and 300,000 authentic, recent, and business-relevant records.
- Use of relative date filters (e.g., last 3, 6, or 12 months) rather than hard-coded dates to ensure ongoing relevance when retraining.
- Iterative training until an acceptable similarity solution is achieved, allowing time for reviewing and updating definitions.
Fields to include in the solution
Customers should select fields that contain meaningful text to identify similarities. The similarity fields must be a subset of the input fields and relevant to the record state (e.g., avoid selecting Close note for open incidents because such text does not exist in open records). These similarity fields are available to users during record creation.
Understanding the similarity score
The similarity score ranges from 0 to 100, indicating how closely two alert records match textually. Alerts scoring above a configurable threshold are returned as similar. Customers can adjust this threshold based on reviewing similarity examples and training progress using the Show training progress feature.
Training progress and timing
Training duration depends on the size and complexity of the dataset; for example, 100,000 records with several hundred classes may take about five hours. The training process automatically includes:
- Fetching and sending the training data to the nearest training service.
- Preparing the data by removing duplicates.
- Training the model using the prepared data.
- Uploading the trained solution back into the system.
Customers can monitor the training progress on the Solutions page to ensure effective model development.
Similarity solutions enable you to use Machine Learning (ML) to compare the text in a resolved alert record to an open alert record to reuse its resolution approach.
Training a similarity solution
To train a similarity solution, you collect words to compile a collection that Machine Learning (ML) can use to compare text in the Short Description, Description, Source, Type, Resource, and Metric Name fields in a resolved alert to see whether the words in the set match words in an open alert. The resolved alert, which is similar to an open alert, provides an example to show how the open alert can be resolved.
- Ensure that the records you train are not too old and that they are relevant to your business needs. Keep the words in the collection current.
- Do not use hard-coded dates as filters because these filters are not updated when you retrain solutions unless you update them manually before every retraining. Instead, use relative date filters, for example, the last 3 months, last 6 months, or last 12 months.
- Perform training as needed until it provides an acceptable similarity solution. This practice provides you time to review and update your solution definition.
Fields to include in the solution
Record the fields that are likely to contain words and phrases that help the system identify similar records for your solution.
The similarity fields that you select should be a subset of your input field selections. For example, if you select fields from incident records that are in Open state, do not select Close note as a similarity field. Because open records do not include Close note fields, the text cannot be similar.
The similarity fields are available to users when they create records.
About the similarity score
The similarity score is a measure from 0-100 of the degree of similarity between two alert records. Alert records that have a similarity score higher than the threshold that you specify is returned by the solution.
Review similarity examples and their scores using the Show training progress feature to determine whether to either increase or decrease the solution threshold. You can change the threshold value in the Threshold for Similarity Score field.
View training solution progress
Training times vary based on the number of records and classes within the training set. The more records and classes you use, the longer the training can take. For example, a data set containing 100,000 records and several hundred classes can take around five hours to complete.
To show the training solution progress, the ML solution automatically performs the following activities when you select Show training progress on the Solutions page. For more information, see View solution training progress.| Activity | Description |
|---|---|
| Fetching files for training. | The system downloads the training records and sends them to the nearest training service. |
| Preparing the data. | The system removes duplicate records from the training set. |
| Training the solution. | The training service trains the solution. |
| Uploading the trained solution. | The training service uploads the solution as attachment records. |