Question

Are there pattern recognition algorithms where I can specify the weight of matching or missing certain parameters? For example, suppose I have 3 strings:

str1 = Samsung 11.6" 64GB Slate PC Tablet with Wi-Fi - Black  
str2 = Samsung Series 7 XE700T1A-A05US 11.6-Inch Slate (64 GB, Win 7 Pro)
str3 = Samsung Series 7 XE700T1A-A03US 11.6-Inch Slate (128 GB SSD, Win 7 HP)

I would like to match str2 to str1 since they have equal GB, even though conventional string distance would say str2 is closer to str3. In reality, I would hope for anything that can handle a large number of parameters with different weights.

Any pointers to the right direction would be appreciated.

Answer 1

1. Create dictionary of named entities, e.g. “Samsung”, “Inch”, “Wi-Fi”, etc.
2. Create set of rules to extract features, e.g. {Number}{Space}?"GB" for gigabytes or {TradeMark}{Space}"Series"{Space}{Number} for trademark and series.
3. Create feature vector, setting known attributes to corresponding values and unknown attributes to something like “NA” (Not Available).
4. Measure distance between items (represented as feature vectors), not strings.

Note: to easier work with dictionaries and rules consider using GATE framework. To measure distance between 2 vectors you may use cosine distance.