Explainable and automated decisions in computer-based reasoning systems
The techniques herein include using an input context to determine a suggested action. One or more explanations may also be determined and returned along with the suggested action. The one or more explanations may include (i) one or more most similar cases to the suggested case (e.g., the case associated with the suggested action) and, optionally, a conviction score for each nearby cases; (ii) action probabilities, (iii) excluding cases and distances, (iv) archetype and/or counterfactual cases for the suggested action; (v) feature residuals; (vi) regional model complexity; (vii) fractional dimensionality; (viii) prediction conviction; (ix) feature prediction contribution; (x) conviction ratio; (xi) contribution ratio; and/or other measures such as the ones discussed herein, including certainty. In some embodiments, the explanation data may be used to determine whether to perform a suggested action.
1. A system including one or more computing devices for performing a machine-executed operation involving instructions, wherein said instructions are instructions which, when executed by the one or more computing devices, cause performance of a process comprising:
receiving a request for a suggested action based on an input context, in a case-based reasoning system, wherein the case-based reasoning system includes a case-based reasoning model;
determining one or more candidate cases based on the input context and the case-based reasoning system, wherein the one or more candidate cases include respective one or more candidate actions;
determining the suggested action based on the respective one or more candidate actions;
determining a regional model of two or more cases in the case-based reasoning model near a suggested case containing the suggested action;
determining a local conviction score for the one or more candidate cases-determining a regional conviction score for the regional model of two or more cases, wherein:
determining the local conviction score for the one or more candidate cases comprises determining the local conviction score as a local familiarity conviction score, wherein the local familiarity conviction score is calculated as a measure of how much information a case represented as a point in the case-based reasoning system distorts the case-based reasoning model; and
determining the regional conviction score for the regional model of two or more cases comprises determining the regional conviction score as a regional familiarity conviction score, wherein the regional familiarity conviction score is calculated as measure of how much information the point distorts the case-based reasoning model;
determining a conviction ratio as a function of the local conviction score and the regional conviction score determining whether to perform the suggested action based on the conviction ratio of the local conviction score and the regional conviction score;
when it is determined to perform the suggested action based on the conviction ratio, causing performance of the suggested action;
when it is determined not to perform the suggested action based on the conviction ratio:
determining one or more explanation factors for the suggested action determined based at least in part on the input context, wherein the one or more explanation factors include at least the conviction ratio determined as a function of the local conviction score and the regional conviction score;
providing at least the one or more explanation factors in response to the request for the suggested action.
2. The system of claim 1 , wherein:
determining the local conviction score for the one or more candidate cases comprises determining the local conviction score as a local prediction conviction score, wherein the local prediction conviction score is calculated as a measure of information required to describe a position of the point in question relative to existing cases represented as points in the case-based reasoning system;
determining the regional conviction score for the regional model of two or more cases comprises determining the regional conviction score as a regional prediction conviction score, wherein the regional prediction conviction score is calculated as the measure of information required to describe the position of the point in question relative to the existing points.
3. The system of claim 1 , wherein the regional model of two or more cases comprises a subset of all the cases in the case-based reasoning model.
4. The system of claim 1 , wherein the input context is a current context for an image labeling system during operation of the image labeling system, and causing performance of the suggested action comprises causing a suggested labeling to be performed by the image labeling system.
5. The system of claim 1 , wherein the input context is a current context for a self-driving car during operation of the self-driving car, and causing performance of the suggested action comprises causing the suggested action to be performed by the self-driving car.
6. The system of claim 1 , wherein the input context is a current context for a manufacturing control system during operation of the manufacturing control system, and causing performance of the suggested action comprises causing the suggested action to be performed by the manufacturing control system.
7. A non-transitory computer readable medium storing instructions which, when executed by one or more computing devices, cause the one or more computing devices to perform a process of:
receiving a request for a suggested action based on an input context, in a case-based reasoning system, wherein the case-based reasoning system includes a case-based reasoning model;
determining one or more candidate cases based on the input context and the case-based reasoning system, wherein the one or more candidate cases include respective one or more candidate actions;
determining the suggested action based on the respective one or more candidate actions;
determining a regional model of two or more cases in the case-based reasoning model near a suggested case containing the suggested action;
determining a local conviction score for the one or more candidate cases;
determining a regional conviction score for the regional model of two or more cases wherein:
determining the local conviction score for the one or more candidate cases comprises determining the local conviction score as a local familiarity conviction score, wherein the local familiarity conviction score is calculated as a measure of how much information a case represented as a point in the case-based reasoning system distorts the case-based reasoning model; and
determining the regional conviction score for the regional model of two or more cases comprises determining the regional conviction score as a regional familiarity conviction score, wherein the regional familiarity conviction score is calculated as measure of how much information the point distorts the case-based reasoning model;
determining a conviction ratio as a function of the local conviction score and the regional conviction score determining whether to perform the suggested action based on the conviction ratio of the local conviction score and the regional conviction score;
when it is determined to perform the suggested action based on the conviction ratio, causing performance of the suggested action;
when it is determined not to perform the suggested action based on the conviction ratio:
determining one or more explanation factors for the suggested action determined based at least in part on the input context, wherein the one or more explanation factors include at least the conviction ratio determined as a function of the local conviction score and the regional conviction score;
providing at least the one or more explanation factors in response to the request for the suggested action.
8. The non-transitory computer readable medium of claim 7 , wherein:
determining the local conviction score for the one or more candidate cases comprises determining the local conviction score as a local prediction conviction score, wherein the local prediction conviction score is calculated as a measure of information required to describe a position of the point in question relative to existing cases represented as points in the case-based reasoning system;
determining the regional conviction score for the regional model of two or more cases comprises determining the regional conviction score as a regional prediction conviction score, wherein the regional prediction conviction score is calculated as the measure of information required to describe the position of the point in question relative to the existing points.
9. The non-transitory computer readable medium of claim 7 , wherein the regional model of two or more cases comprises a subset of all the cases in the case-based reasoning model.
10. The non-transitory computer readable medium of claim 7 , wherein the input context is a current context for an image labeling system during operation of the image labeling system, and causing performance of the suggested action comprises causing a suggested labeling to be performed by the image labeling system.