In one embodiment, a learning machine device initializes thresholds of a data representation of one or more data features, the thresholds specifying a first number of pre-defined bins (e.g., uniform and equidistant bins). Next, adjacent bins of the pre-defined bins having substantially similar weights may be reciprocally merged, the merging resulting in a second number of refined bins that is less than the first number. Notably, while merging, the device also learns weights of a linear decision rule associated with the one or more data features. Accordingly, a data-driven representation for a data-driven classifier may be established based on the refined bins and learned weights.