# Measure similarity between $2$ variable length sequences using LSTM

I would like to build a model that measures the similarity between $2$ sequences that have different lengths.

Input: $[a_1, a_2, \dots, a_{n-1}, a_n]$ and $[b_1, b_2, \dots, b_{m-1}, b_m]$ where $a_i$ and $b_i$ are sets of features.

Output: score $s \in \mathbb{R}$

More precisely, I want to measure the similarity between $2$ signatures, where a signature is defined as a sequence of features, measured while the signature is being handwritten, every $\Delta t$ seconds. The set of features comprises timestamp, azimuth, altitude, pressure, $(x,y)$ coordinates on the tablet ...

My first idea is to feed a LSTM with $n$ data points (data point $=$ features at given time $t$) of each sequence; which is $2n$ data points because we are comparing $2$ sequences of data points.

1. Does this sound like something that could work?
2. I have a dataset collected from $5$ users. For each user, I have 20 genuine signatures and 20 forged signatures. Is this enough?
• You should spell out LSTM, what does it mean? – kjetil b halvorsen Sep 4 '16 at 11:41
• Long short-term memory, google it :) – tgy Sep 5 '16 at 10:49